International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No.

6, 2024

Breast Cancer Diagnosis with Machine Learning

Using Feed-Forward Multilayer Perceptron
Analog Artificial Neural Network
1
Djimeli-Tsajio Alain B. , Koagne Longpa T. Silas 2,*, Lienou T. Jean-Pierre3, Noulamo Thierry3,
and Geh Wilson Ejuh4,5
1
Department of Telecom. and Network Engineering, IUT-FV, University of Dschang, Bandjoun, Cameroon
2
Research Unit of Automation and Applied Computer Science URAIA, IUT-FV,
University of Dschang, Bandjoun, Cameroon
3
Department of Computer Engineering, IUT-FV, University of Dschang, Bandjoun, Cameroon
4
Department of General and Scientific Studies, IUT-FV, University of Dschang, Bandjoun, Cameroon
5
Department of Electri. and Electro. Eng., NAHPI, University of Bamenda, Bambili, Cameroon
Email: [email protected] (D.T.A.B.), [email protected] (K.L.T.S.),
[email protected] (L.T.J.-P.), [email protected] (N.T.), [email protected] (G.W.E.)

Abstract—An analog network classifier based on a research on human brain functioning, ANNs are capable
multiplier and non-linear functions is presented in this of learning from experience.
paper, executing binary classification on breast cancer cells, An ANN is defined as an information processing
and categorizing biopsies as benign or malignant tumors.
An off-chip learning on-chip inference methodology is system inspired by the functioning of the human brain to
proposed for implementing a feed-forward analog artificial learn [6]. On the other hand, an analog implementation
neural network based on fundamental design analog block simply refers to the design and realization of a system
circuits, realized with the aid of 90 nm CMOS technology. using electronic components. Thus, the implementation
These circuits are meticulously designed and fine-tuned at of an Analog Artificial Neural Network (AANN) is then
the transistor scale to meet design criteria while minimizing referred to as the design and realization of an information
power consumption. Through Spice simulations, the basic
analog blocks were developed, leading to the specification of
processing system inspired by the functioning of the
the full-chip hardware neural network. The Monte Carlo human brain to learn using electronic components.
analysis of the final circuit reveals that the network achieves ANNs are particularly attractive for Complementary
96.85% accuracy and 0.9309 MCC on the Wisconsin breast Metal-Oxide Semiconductor (CMOS) Very Large-Scale
cancer dataset, with a power consumption of 31.95 µW, and Integration (VLSI) implementations because each parallel
power supply rail of ±900 mV per analog circuit component element (that is, neuron or synapse) is relatively realized
and computational unit. The model effectively captures data with the aid of electronic components, allowing the
patterns, providing stable, reliable, and robust predictions.
complete integration of large networks on a single chip [7,
Index Terms—Analog artificial neural network, very large- 8]. Multipliers, nonlinear functions (also called activation
scale integration, complementary metal-oxide functions), and their derivatives are key elements in an
semiconductor, breast cancer classification, multilayer Analog Very Large-Scale Integration (AVLSI)
perceptron implementation of ANNs [9].
Breast cancer cells classification as malignant or
I. INTRODUCTION benign tumors was chosen for the verification of this
work. Hence, it will be necessary to find the best
Breast cancer remains one of the most common compromises of obtained analog structures through
cancers and the leading cause of mortality among women simulations in Cadence Virtuoso for achieving optimized
worldwide. This spotlights the urgent need for early and circuits in mature 90 nm CMOS technology. The ANN
efficient diagnostic methods. Over the past ten years, the studied in this work is based on the Multilayer Perceptron
use of Artificial Neural Networks (ANN) has developed (MLP) architecture, whose learning phase uses the back-
in many disciplines [1]. They are particularly applied for propagation technique [10–12]. To validate this work,
classification, prediction, optimization, and associative simulations will be made on the structure of the obtained
memory [2–4]. ANN resulted from the admiration of how analog breast cancer classifier using the above-mentioned
the human brain computes complex processes, which is technology in order to compare the obtained results with
entirely different from the way conventional digital those found in the state of the art [7, 13–18].
computers do this [5]. Drawing their inspiration from Within this work, the use of feed-forward architecture
is fundamental. Information flows in a unidirectional
Manuscript received April 25, 2024; revised June 6, 2024; accepted manner from input to output layers [19], allowing the
June 12, 2024.
*
Corresponding author
model to capture, and analyze complex patterns and

doi: 10.18178/ijeetc.13.6.427-441 427

 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

features relevant to breast cancer detection. The MLP produce the desired input and output signals used to test
structure, known for its capability to handle intricate data breast cancer Integrated Circuits (ICs) [8]. These signals
relationships [20], is strategically applied to address have amplitudes of 1 V and operate at frequencies up to 5
challenges and opportunities inherent in precise breast kHz. In 2018, Hassan Jouni et al. designed a multiplier,
cancer classification using analog neural networks. an activation function, and its derivative used in the
Most ANNs used for tasks such as pattern recognition realization of a feed-forward ANN based on the MLP
and medical diagnosis are digital, meaning they rely on with a backpropagation algorithm under MATLAB
discrete data and numerical precision [20]. In contrast, Simulink [9]. In 2019, Hassan Jouni et al. implemented in
AANN utilizes continuous values, offering advantages in MATLAB Simulink an AANN for the detection and
modeling biological phenomena and handling subtle classification of breast cancer based on non-ideal blocks
variations. AANN excel due to their robustness against with a classification error of 2.6% [10, 11].
noise and parallel processing capabilities. Additionally, That same year, Akshay Jayaraj et al. designed a
they are more energy-efficient and adaptable, offering common source nonlinear function that consists of an
high-speed computation and chip dimension reduction, NMOS transistor coupled to a PMOS transistor [13]. To
aligning better with biological principles. Using AANN this day, this nonlinear function is the most commonly
for medical diagnostics, especially in breast cancer used activation function in AANN. In 2020, Sanjeev T. et
detection context, is relatively uncommon, making the al. proposed a fully integrated AANN based on the
research innovative and pioneering. This approach opens common source amplifier for breast cancer classification
up new possibilities for achieving real-time diagnosis. through a hardware-software co-design methodology
Furthermore, the primary motive behind this research with about 96.59% accuracy operating at 1.1 V [15]. In
work is to enhance the early detection of breast cancer 2021, Alejandro Medina-Santiago et al. proposed a
through the development of advanced computational CMOS implementation of an AANN based on the
methodologies. By leveraging an AANN tailored optimization of an N-dimensional objective function
specifically for breast cancer detection, the research aims having a 5-3-1 configuration architecture with a total
to achieve the ability to identify subtle indicators of power consumption of 46.08 mW [14]. In the same light,
malignancy at an earlier stage, potentially leading to Gencer et al. proposed a paper focusing on the design and
improved patient outcomes and survival rates. The validation of an AANN [22]. In their work, basic building
application of analog computing principles to the domain blocks of the AANN were constructed in UMC 90 nm
of medical analysis, particularly for breast cancer technology, revealing that the AANN operates with a
detection, seeks to optimize computational resources and performance accuracy of 99.85% on the XOR gate.
potentially enable real-time diagnosis, which is critical Additionally, Freeman et al. [23] reviewed the use of
for improving the accessibility of diagnostic procedures. ANNs for detecting breast cancer across various imaging
The methods presented in some previous studies techniques, highlighting accuracy improvements and
reported in literature have the following limitations: fixed screening advancements. Mina et al. [24] reviewed
weights without on-chip learning and use of a common machine learning techniques, shedding light on
source amplifier as a non-linear function; use of a Metal automation strategies particularly relevant to circuit
Oxide Semiconductor Field-Effect Transistor (MOSFET) design. Houssein et al. [25] proposed an optimized deep-
operating only in weak inversion, resulting in low time learning architecture for breast cancer diagnosis,
response; use of unconventional architecture; low noise integrating an improved marine predator algorithm to
immunity; low speed; high complexity of the proposed enhance performance. Concurrently, Ye et al. [26]
systems; and complex communication between modules introduced a method of analog deep neural network noise
with respect to the operations performed by each circuit. injection, ensuring stable prediction results despite
This paper is organized as follows: a brief literature parameter perturbations, which is vital for safety-critical
review is presented in Section II, followed by the applications. Ben Ahmed et al. [27] presented a hybrid
mathematical background in Section III. Section IV UNET model for early breast cancer detection through
highlights the AANN circuit design methodology used. ultrasound image segmentation, showcasing novel
Simulation results are presented in Section V. Section VI approaches to diagnosis. Lastly, Moges et al. [28]
then focuses on the interpretation and validation of the conducted a comparative analysis of ANN and support
obtained results. Finally, a brief conclusion and vector regression (SVR) models in predicting spray drift,
perspective for further studies are given in Section VII. contributing to environmental science research.
Sedlakova et al. [29] suggested that digital methods in
II. LITERATURE REVIEW medical diagnostics enhance accuracy, but present
complexity and high-cost challenges. In contrast, analog
Studies in this area began to take shape in 2016, and a techniques are simpler, cost-effective and more reliable,
brief review will help facilitate an in-depth understanding yet face stability challenges. This work however captures
of different methods deployed within this work. In 2016, complex non-linear relationships between input and
Jouni et al. designed a low-complexity architecture for output variables, enabling highly complex system
breast cancer classification by pattern recognition [21]. modeling. Thereby, significantly improving accuracy,
They also focused on determining the activation function energy efficiency, and speed processing, making it
that minimizes classification errors. In 2017, Hassan worthy for breast cancer detection. Thus, potentially
Jouni et al. designed a programmable signal generator to revolutionizing the field of medical diagnostics.

428
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

III. MATHEMATICAL BACKGROUND dependents on both VX and VY. Hence, VX and VY must be
kept small to maintain good linearity. This Gilbert cell
Gilbert cells are essential components in the structure will be used in this work for basic circuitries.
architecture of AANNs due to their intrinsic ability to
enable analog signal processing without the need for IV. DESIGN OF THE CIRCUIT BLOCKS
digitization [30]. This is achieved without necessitating
the conversion of these signals into a digital format. The ANNs are most efficiently implemented using analog
operational principle of Gilbert cells involves the circuits [33]. Analog implementations are generally faster
generation of an output signal directly proportional to the and require less hardware (fewer transistors) than digital
product of two distinct input signals [31]. VLSI implementations [4]. AVLSI implementations are
This cell is depicted in Fig. 1 [32], consisting of three parallel systems used in solving real-world problems [14],
differential pairs (M1-M2, M3-M4, and M5-M6). The [34]. Multipliers and non-linear functions based on
inputs: VX+ is applied to the gates of M2 and M3, VX- is Gilbert cells are key elements in performing analog
applied to the gates of M1 and M4, VY+ is applied to the calculations efficiently [35]. Based on MLP architecture,
gates of M6, and VY- is applied to the gates of M5 each synapse needs a multiplier to multiply each input by
respectively. The outputs: (IO+ and IO-) are taken from the the corresponding weight, and each neuron needs a
collector pairs (M1, M3) for I1, and (M2, M4) for I2 nonlinear function and its derivative [7].
respectively. Assuming all the transistors are biased in The feed-forward multipliers derived from the Gilbert
cell will multiply the input, VX by each corresponding
the saturation region, obeying the ideal square-law
synaptic weight, VW to produce an output current, IO as
equation, sized and matched so that the transconductance
previously mentioned above. Unlike some neural
parameters satisfy k1 =k2 =k3 =k4 =ka and k5 =k6 =kb . networks that require an explicit adder block to combine
weighted inputs, the Gilbert cell structure obviates this
need. The current outputs from the hidden layer naturally
sum up due to Kirchhoff’s Current Law (KCL). Then the
non-linear neural function will take the sum of all
synapse multipliers connected to a neuron and produce an
output. The output of the non-linear function will be the
input of the derivative block for calculating the derivative
of the non-linear function. As an example, the non-linear
function can be the sigmoid (or logsigmoid) function [7].
It should be noted that an AANN system can contain
thousands of multipliers and non-linear functions [36, 37].
It is crucial to optimize these analog circuit blocks to
ensure precision, efficiency, and scalability. As a result,
the surface area and the energy consumption of each
block must be as low as possible. Based on the above
Fig. 1. MOS version of Gilbert cells [32]. statement, the methodology and design techniques of this
work are initially based on the following six steps:
Defining the output current IO of Fig. 1 as Step 1: Choosing the best circuit topology suited to our
I O =I 2  I1 =   I 2b  I 2a    I1a  I1b  , it can be shown that: algorithm with respect to the state of the art.
Step 2: Dimensioning each basic block in order to
 k V2 k V2  optimize the specifications in terms of dynamics and
I O  2ka VX  I 3 1  a X  I 4 1  a X  . If we
 2I3 2I4  precision.
  Step 3: Choosing a low-cost and mature CMOS VLSI
2 2
kV kV technology to achieve optimization at the transistor level.
demand a X  1 and a X  1 , it follows that IO Step 4: Using specialized and dedicated software to
2I3 2I 4
simulate and validate the main characteristics of the
depends linearly on VX: proposed circuits.
IO  2ka  I3  I 4 VX Step 5: Implementing a feed-forward AANN based on
MLP architecture using an off-chip learning on-chip
inference methodology.
While the currents I 3 , I 4 can be expressed by VY: Step 6: Simulating and testing the AANN performance
on breast cancer classification in Cadence Virtuoso.
VY =
1
kb
 I3  I 4  Incorporating an off-chip learning on-chip inference
method enhances flexibility, performance, and ensures a
reliable AANN implementation for medical diagnostics.
Substituting VY into the expression of IO, it follows that:
A. Multiplier Design
I O  2ka kb VX VY (1) Multipliers are essential elements in analog signal
processing, particularly in AVLSI implementations [38].
The output current in (1) yields an ideal analog The design of this multiplier must take into account
multiplier. Notice that since both I3, I4 are ISS and IO transistor biasing, linear or saturation regime, weak or

429
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

strong inversion, and threshold [7]. This section focuses inversion remain complex, but can be approximated in
on the design of a four-quadrant multiplier. terms of voltages or currents:
The main characteristics of a multiplier are its input  VGS > VT + 100 mV for strong inversion;
and output dynamics, linearity, and bandwidth [39]. The  VT + 100 mV > VGS > VT –100 mV for moderate
design of an analog CMOS multiplier with high input inversion;
dynamics and very low power consumption is proposed.  VGS < VT –100 mV for weak inversion.
One of the input voltages of this multiplier corresponds to It is often preferable to design a circuit by defining the
the weight, VW, used in the ANN algorithm for breast bias current. The following definitions can then be used
cancer detection and classification, meanwhile, the other to frame the limits of inversion:
input, Vin, corresponds to the biopsy cancer input [21].  IO > 10IS for strong inversion;
The inputs of each multiplier are then chosen as voltages
 10IO > IO > 0.1IS for moderate inversion;
and the outputs as currents. The adder inputs will
 IO < 0.1IS for weak inversion.
therefore be currents. This technique is deployed to
eliminate the use of an analog adder in feed-forward where IS corresponds to the moderate reversal
propagation, thereby reducing the complexity of the characteristic current and is defined by Eq. (2), in which
overall network. W and L represent respectively the width and length of
Moreover, the 9 attributes of the 699 biopsies each transistor in Fig. 2.
correspond to voltages between 30 mV and 300 mV with 2 COX 2 W 2 K P 2 W
a scale of 10 mV per coefficient (weight or input) [7]. IS  UT  UT (2)
K L K L
Concerning the weight values, which are respectively
between −1.37 and 1.17 in the hidden layer and between with K P and COX given by (3), and (4) respectively.
−3.54 and 3.48 in the output layer [21], the input dynamic
of the multiplier is then imposed by the weights, which K P   COX (3)
are higher than those of the attributes. The minimum and
 0 r  OX
maximum weight values in the hidden and output layers COX   (4)
vary between −3.54 and 3.48 for the logsigmoid TOX TOX
activation function. The input dynamics of the analog
CMOS multiplier are therefore −354 mV to 348 mV. With UT=26 mV representing the thermal voltage, and
Taking into account the input noise effects, this dynamic defined by: UT=KT/q, K=0.7,  r = 3.9,  0 = 8.854×10-12
should increase up to 380 mV (or 400 mV). Besides, by and μ represents the charge mobility.
applying a multiplying factor of 0.1 to the attributes Table I summarizes the Spice parameter characteristics
instead of 0.3 as before, the input voltage dynamics of the of the transistors (nmos1v8 and pmos1v8) chosen from
analog multiplier will now correspond to: HCMOS9, a 90 nm CMOS technology from
 10 mV to 100 mV for the first input, Vin and STMicroelectronics. The parameters provided include
 ±400 mV (or −354 mV to 348 mV without noise) for specifications on oxide thickness, oxide capacitance,
the second input, VW. charge mobility, and transconductance, essential for
circuit simulation and analysis. Recalling permittivity:

 OX   0 r  3.9  8.854 1012  3.45 1011 F/m

TABLE I: TRANSISTOR SPICE PARAMETERS
Spice parameters NMOS PMOS
TOX (nm) 2.05 2.15
COX (F/m2) 1.68×10-2 1.60×10-2
µ(m2/Vs) 2.98×10-2 6.27×10-2
KP (µA/V2) 500.64 100.32

To limit the final consumption of the complete

multiplier, a polarization current of the complete structure
(presented in Fig. 2) is fixed at IO=1 µA. For optimal
operation of the corresponding transistors (differential
Fig. 2. Multiplier circuit. pairs) in a strong inversion regime, the IO>10IS condition
must be verified.
Taking into account the considerations presented in the
 For NMOS transistors, IS can be fixed to:
preceding paragraphs, we have the four-quadrant
ISN=IO/10=1 μA/10=0.1 μA.
multiplier in Fig. 2.
Considering Eq. (2), we obtained for the NMOS
To explain the behavior of the 4-quadrant multiplier,
transistors:
let us recall some transistor operation characteristics. A
transistor does not move from its exponential behavior in 2 K PN 2 WN
ISN  UT (5)
weak inversion to its quadratic behavior in strong K LN
inversion; the transition is understood to be hybrid [7].
WN
The boundaries between weak, moderate, and strong  0.1034 (6)
LN

430
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

 For PMOS transistors, IS can be fixed to:

ISP=IO/20=1 μA/20=0.05 μA.
Considering Eq. (2), we obtained for the PMOS
transistors:
2 K PP 2 WP
ISP  UT (7)
K LP

WP
 0.2580 (8) Fig. 3. Linear current to voltage converter circuit.
LP
Taking W1=W2=W and L1=L2=L, we have the following:
By taking into consideration the multiplier circuit
presented in Fig. 2 and choosing WN=120 nm, we obtain 1 W
LN=1.16 µm for transistors NM4, NM5, NM11, and
  KPN (9)
2 L
NM12. By setting WP=120 nm for PMOS transistors, we
deduce LP=465 nm for the transistors PM13 and PM14. VGS1=VOUTVSS and VGS2=VDDVOUT
The transistor sizing current generator is based on the
I1   VOUT  VSS  VTH 
2
(10)
following relationship: I O  Vin I1 , where I1 follows a
I 2   VDD  VOUT  VTH 
quadratic law according to the weight input, VW. To size 2
(11)
the transistors, we considered the maximum values of the
inputs. Table II summarizes the dimensioning of the From KCL, IIN=I1-I2, so IIN can be written by taking
transistors in the complete multiplier circuit. the difference between (10) and (11) as in (12).
TABLE II: DIMENSIONS OF THE TRANSISTORS FOR THE MULTIPLIER I IN   VDD  VSS  2VTH  2VOUT  VSS  VDD  (12)
W/L W/L
Transistor Type Transistor Type
(nm/µm) (nm/µm) Knowing that VSS=VDD, we deduce the output voltage,
NM1 NMOS 150/10.0 PM8 PMOS 640/0.9
NM2 NMOS 150/1.55 NM9 NMOS 500/1.0 VOUT as a function of the input current IIN as in (13).
NM3 NMOS 580/6.0 NM10 NMOS 150/0.15
NM4 NMOS 120/1.16 NM11 NMOS 120/1.16 I IN
VOUT  (13)
NM5
PM6
NMOS
PMOS
120/1.16
180/3.0
NM12
PM13
NMOS
PMOS
120/1.16
120/0.465
4 VDD  VTH 
PM7 PMOS 500/0.51 PM14 PMOS 120/0.465
We obtained the width and length (W/L) of NM1 and
NM2 in Fig. 3 to be 150 nm and 950 nm, respectively.
B. Linear Current to Voltage Converter Design
C. Logsigmoid Type Activation Function Design
Before defining the AANN topology, the design of a
linear current-to-voltage converter is needed. This section Feed-forward network responses are usually nonlinear
is dedicated to this purpose with transistors biased in functions that are achieved by summing the output
weak inversion. This method makes it possible to obtain a currents of the synapses at each node. These functions are
significant gain with low power consumption, and easily realized by the activation function circuit. More still, the
realize a logsigmoid function between the output of the derivative of the nonlinear function with respect to the
multiplier and the input of the chosen activation function. weighted input is required for a fully analog on-chip
K. Bult and H. Wallinga [40] proposed a linear learning neural using the backpropagation algorithm [37].
current-to-voltage converter that has only two transistors. In this work, we utilize the sigmoid activation function
The two NMOS transistors, NM1 and NM2, both [12], which consists of two linear current-to-voltage (IV)
functioning in a biased saturation regime. To eliminate converters and a nonlinear function circuit. These IV
the body effect on the threshold voltage, VTH, the gate of converters transform input and output currents into
the two transistors is connected to their sources, as shown corresponding input and output voltages as depicted in
in Fig. 3. Fig. 4 and governed by Eq. (14).

Fig. 4. Logsigmoid type activation function circuit.

431
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

1 calculated by the numerical finite difference method.

Vout  (14) After finding the change in each weight, the program
1  eVin 
  updates the weight file. Subsequently, the circuit
  simulation is performed again with the newly calculated
where Vin represents the input voltage of the logsigmoid weights, as illustrated in Fig. 5. The flowchart loop runs
activation function circuit obtained at the output of the until the tolerance criterion is reached, or the maximum
linear current to voltage converter located at the input of iteration number is exceeded.
Fig. 4.
I IN
VOUT(IV-converter)  Vin(logsig) 
4 VDD  VTH 
This function is equally approximated using transistors
and has the same working principle as that of the
hyperbolic tangent activation function [7]. Table III
summarizes the dimensions of the transistors used in the
circuit shown in Fig. 4.
TABLE III: DIMENSIONS OF THE TRANSISTORS FOR THE LOGSIGMOID
TYPE ACTIVATION FUNCTION
W/L W/L
Transistor Type Transistor Type
(nm/µm) (nm/µm)
NM1 NMOS 240/5.4 NM11 NMOS 150/1.5
NM2 NMOS 275/6.2 NM12 NMOS 150/1.5
NM3 NMOS 250/6.0 PM13 PMOS 1000/1.0
NM4 NMOS 280/5.1 PM14 PMOS 1000/1.1
PM5 PMOS 300/2.0 NM15 NMOS 150/1.5
PM6 PMOS 300/2.0 NM16 NMOS 150/1.1
PM7 PMOS 600/1.6 NM17 NMOS 150/1.1
PM8 PMOS 1000/1.0 NM18 NMOS 230/6.0
PM9 PMOS 1000/1.1 NM19 NMOS 150/10.0
NM10 NMOS 150/1.5 NM20 NMOS 150/10.0

D. Feed-Forward Analog Neural Network Layers Design Fig. 5. 9-10-2 analog network training flowchart.
In the literature, several ANN learning techniques have In addition, Fig. 6 shows the architectural layout of the
been proposed to find an accurate weight set [41–43]. In proposed fully integrated off-chip learning on-chip
this work, the gradient descent backpropagation inference feed-forward AANN implemented using analog
algorithm is selected to train the neural network [44, 45]. circuits comprising logsigmoid activation functions and
Since the proposed ANN topology has no constraints on four-quadrant multipliers designed in 90 nm CMOS
the learning performance, the chosen algorithm provides technology. This CMOS implementation can be
a simple and efficient implementation. The algorithm is considered as multi-layer, fully connected AANN that
developed in the MATLAB environment, and the input consists of an input layer, a hidden layer, and an output
parameters are extracted from circuit simulations of the layer, as modeled by (15) and (16). The input layer of
designed AANN. The system is fully automated, and all this breast cancer classifier comprises nine inputs
simulations can be performed in Cadence. Fig. 5 shows corresponding to the attributes of the database. The
the implemented process steps. hidden layer contains ten neurons, which apply nonlinear
Furthermore, the procedure starts by creating weight transformations to the inputs. The output layer consists of
files for the ANN system to operate correctly. Afterward, two neurons, which correspond to the output of the
a random set of weights is sent to the circuit simulator to network. The outputs are classified as either benign (class
find the output values corresponding to that weight set. 1: 01 in binary) or malignant (class 2: 10 in binary).
Simulation outcomes are then transmitted to the
algorithm, which then calculates the error between the  9 
actual sample outputs and the AANN results. a j  g j   wi , j xi  ; j  1, 2,...,10 (15)
Also, the stochastic gradient descent method is used to  i 1 
compute the derivative of the mean squared error (SSE)
 10 
function, which is used as the cost function [46, 47]. If all sk  g k   w j , k a j  ; k  1, 2 (16)
 
outputs produced by the AANN for each sample satisfy  j 1 
the desired tolerance criteria, the program stops the
learning process with the last weight set. If the error is where: xi are the inputs (the 9 cancer attributes of the
greater than the tolerance, the backpropagation algorithm Wisconsin Data Base), wi,j represents the 90 weight
attempts to minimize SSE. In particular, the change in coefficients in the hidden layer, aj corresponds to the 10
each weight is estimated by calculating the derivative of outputs of the hidden layer, wi,j represents the 20 weight
the cost function for the corresponding weight. coefficients in the output layer, and sk corresponds to the
Consequently, the derivative of the logsigmoid plays a 2 neurons at the outputs of the network. The activation
central role. The derivative of the activation function function (gj and gk) of the layers is the logsigmoid, as
from the output curve of the logsigmoid circuit is earlier mentioned.

432
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

Fig. 6. 9-10-2 analog network circuit block diagram.

Off-chip learning is utilized in this work for several −360 mV to 360 mV, which equally respects the weight
reasons. To begin with, it allows rapid training processes. specifications. In conclusion, this second multiplier input
Additionally, it saves much silicon area as the circuits can be used as the weighted input for breast cancer
dedicated to learning are not needed [48]. Moreover, off- applications.
chip learning works with chip-in-the-loop learning, which
helps to address the mismatch between the network’s
behavioral model and the analog implemented circuit. As
a result, any flaws in the hardware chip can be accounted
for during the training stage, allowing for the updating of
weights accordingly. This work is designed to facilitate
efficient signal processing, learning, and decision-making
using CMOS circuitries sized and optimized at the
transistor levels for breast cancer diagnosis.

V. SIMULATION RESULTS
A. Multiplier
1) Direct Current (DC) response analysis
The designed multiplier offers a linear output current
Iout with respect to the input voltage Vin ranging from Fig. 7. Iout in function of Vin for different values of VW.
−360 mV to 360 mV. Thus, respecting the Wisconsin
cancer input attribute specification dynamics as earlier
mentioned, which range from −100 mV to 100 mV. This
input corresponds to the nine attributes of the Wisconsin
database, whose values are from 1 to 10, with 1
corresponding to a voltage of 10 mV and 10
corresponding to a voltage of 100 mV. This first input
dynamic is sufficient for breast cancer applications, and
its DC simulation response is shown in Fig. 7.
The second input of the designed multiplier
corresponds to the weights of each neuron. This value
must be converted as a function of the square of the drain
current to have a linear relationship with the output of the
multiplier. The corresponding DC simulation response is
shown in Fig. 8. This result shows a linearity range from Fig. 8. Iout in function of VW for different values of Vin.

433
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

Fig. 7 and Fig. 8 allow validation of the input

dynamics considerations of the multiplier in DC domain
used for AANN dedicated to the classification of breast
cancer cells.
2) Transient response analysis
The transient response is presented for two different
configurations with a maximum dynamic input voltage.
In the first configuration shown in Fig. 9, the cancer input
attribute voltage, Vin represented by the green sine wave,
has a frequency of 100 Hz with an amplitude of 100 mV,
and the weight input, VW represented by the blue sine
wave, has a frequency of 1 kHz with an amplitude of 1 V,
producing an output amplitude modulated current signal
represented in red.
In the second configuration shown in Fig. 10, the (a) (b)
frequencies of both inputs are interchanged (that is, 1 kHz
Fig. 11. Iout Spectra: configuration 1 (in green) and configuration 2 (in
for Vin and 100 Hz for VW) which equally produces an blue): (a) Iout spectrum for Vin at 100 Hz and VW at 1 kHz and (b) Iout
output amplitude modulated current signal of the same spectrum for Vin at 1 kHz and VW at 100 Hz.
amplitude and frequency as in Fig. 9.
The spectrum of the output amplitude modulated The various extensive transient simulation results
current signals, Iout is calculated to confirm the transient conclusively demonstrate that the power consumption of
analysis of the multiplier for the two configurations. The the multiplier remains comfortably well below the
results for the two configurations are substantially threshold, with a maximum value not surpassing 3 µW
identical and presented in Fig.11. (actually about 2.9 µW). This efficiency ensures that the
As expected, we obtain two lines at 0.9 kHz and 1.1 multiplier operates within the desired power constraints.
kHz which correspond to the mixtures of the two input
signals at 100 Hz and 1 kHz. Their amplitude in both B. Linear Current to Voltage converter
cases is about −130 dB. Harmonics due to non-linearities 1) DC response analysis
of the multiplier remain below −160 dB, which gives a To be compatible with the multiplier results presented
noise margin of about 30 dB. This value is again largely in the previous section, the input dynamic IIN is
sufficient for breast cancer applications. dimensioned in the range of [−5, +5] µA which is
actually enough for the breast cancer diagnostic
specifications as presented in Fig. 12 (a).

(a)

Fig. 9. Transient output Iout (weighted output in red) for Vin (input of a
layer in green) at 100 Hz and VW (weight coefficient in blue) at 1 kHz.

(b)
Fig. 12. Linear current to voltage converter DC response: (a) Vout in
function of IIN (in green) and (b) Error estimation between Vout and IIN
(in red).

This converter produces an output (colored in blue)

that is directly proportional to the applied input, resulting
in a variation of the output voltage, Vout, in the range
[−100, +100] mV, with the error estimation between the
input and output shown in red in Fig. 12 (b).
2) Transient response analysis
Fig. 10. Transient output Iout (weighted output in red) for Vin (input of a This section focuses on the transient response
layer in green) at 1 kHz and VW (weight coefficient in blue) at 100 Hz. verification of the designed linear current-to-voltage

434
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

converter, which permits the validation of its study with The input dynamic is strictly respected since saturation
respect to the feed-forward AANN conceptual framework of the output voltage is obtained at 0 mV and 100 mV
as shown in Fig. 13 (a). before the extreme input current values. For Vb = 0.6 V,
The input current is illustrated by a sine wave of saturation is reached at IIN = ±2 µA, as seen in Fig. 14 (b).
amplitude 5 µA and frequency 1 kHz which produces a The asymmetry observed corresponds to a value of 0 mV
sine wave output voltage of amplitude 100 mV and at IIN = −2 µA, and 100 mV at IIN = +2 µA, as required.
frequency 1 kHz with no phase shift as depicted in Fig. This leads to the validation of the conceptual framework
13 (c). This leads to the validation of the proposed linear of the designed logsigmoid activation function in the DC
current-to-voltage converter since the spectrums of the domain.
input and output signals respectively shown in Fig. 13 (b)
2) Transient response analysis
and Fig. 13 (d) have similar shapes, with no time shift.
The transient response of the logsigmoid-type
activation function is presented in Fig. 15 (a). The input
current is represented by a sine wave of amplitude 5 µA
and a frequency of 1 kHz, which produces a square wave
output voltage of amplitude 100 mV (which is the ideal
amplitude value) and a fundamental frequency of 1 kHz
with no phase shift, as illustrated in Fig. 15 (c). The
spectrums of the input and output signals respectively
shown in Fig. 15 (b) and Fig. 15 (d) have similar shapes.

(c) (d)
Fig. 13. Linear current to voltage converter transient and spectrum
responses: (a) Linear current to voltage converter transient input signal,
(b) Linear current to voltage converter input signal spectrum of (a), (c)
Linear current to voltage converter transient response, and (d) Linear
current to voltage converter spectrum response of (c).

C. Logsigmoid Activation Function

1) DC response analysis
The output dynamics of the logsigmoid activation (c) (d)
Fig. 15. Logsigmoid type activation function transient response: (a)
function are shown in Fig. 14. It is observed from the Logsigmoid transient input signal, (b) Logsigmoid input signal
figure that the output voltage varies from 0 to 100 mV. spectrum of (a), (c) Logsigmoid transient response, and (d) Logsigmoid
The first configuration (Fig. 14 (a)) shows the output spectrum response of (c).
response for different biasing voltages, Vb ranging from 0
to 1 V with a step of 0.1 V. This output voltage will be D. Breast Cancer Software Classification
used as the input voltage to the nonlinear derivative This work uses the supervised learning classification
circuit. approach. Two types of neural networks are used in this
work to classify cancer tumors, namely, feed-forward and
backpropagation neural networks [49]. The classification
process is divided into a training phase and a test phase.
The neural network trains through the adjustment of the
weights to predict the correct class. The desired output is
specified as class one (01) for benign and class two (10)
for malignant. This process is done through the use of the
backpropagation algorithm. After the training phase is
completed, the stored weights are then used in a feed-
forward architecture for the neural network test phase.
The primary objective of this section is to provide the
results of breast cancer training. MATLAB serves as a
powerful and functional software environment for
constructing or facilitates the creation and simulation of
(a) (b) neural networks through coding [50]. Specifically, we
Fig. 14. Logsigmoid type activation function DC response: (a) DC response focus on training our neural network for breast cancer
for Vb ranging from 0 to 1 V and (b) DC response for Vb at 0.6 V. detection. The algorithm parameters include a learning

435
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

rate of 0.295 and 1000 iterations. Fig. 16 (a) and Fig. 16

(b) show the performance and error measures,
respectively. It is observed from Fig. 16 that the training
of the network runs until the tolerance criterion is reached
at the maximum iteration number. This criterion helps
prevent overfitting by stopping the training process
before the model becomes too tailored to the data.

(a) (b)
Fig. 17. Feed-forward AANN DC response for ten benign samples: (a)
First neuron output and (b) Second neuron output.

(a) (b)
Fig. 16. Neural network test phase software simulation curves: (a)
Neural network performance curve and (b) Neural network error curve.

E. Monte Carlo and Process Corner simulation

AANNs are subject to process variations that vary the
values and the characteristics of the circuit components.
Small deviations may result in acquiring erroneous
outputs, thereby leading to a larger error [51]. Taking into (a) (b)
account the impact of these variations, the designed Fig. 18. Feed forward AANN DC response for ten malignant samples:
AANN is analyzed through Monte Carlo sensitive (a) First neuron output and (b) Second neuron output.
analysis and process corner variations. To perform the
process corner analysis, we simulated five different Similarly, Fig. 19 illustrates the DC response results of
corners based on the manufacturing variations of NMOS twenty close neuron outputs, of which ten are classified
and PMOS devices: typical (tt), fast-fast (ff), slow-slow as benign-class tumors. That is, Fig. 19 (a) and Fig. 19
(ss), slow-fast (snfp), and fast-slow (fnsp). Fig. 17 to Fig. (b). The other ten are classified as malignant-class tumors.
19 show the histograms of logic 1 (in red) and logic 0 (in That is, Fig. 19 (c) and Fig. 19 (d). Despite the similarity
blue) values obtained from the Monte Carlo simulation of of the two neuron outputs from these twenty biopsies, the
the cancer attribute samples for each corner. classification task is successfully accomplished. This
The attributes from the Wisconsin database are demonstrates the reliability of the design analog network
translated in the form of voltages between 10 mV and as an ANN system for classification purposes.
100 mV to perform the simulation in Cadence Virtuoso.
These voltages are used as inputs to the designed AANN
for simulation of the 699 biopsies.
Fig. 17 illustrates the DC response simulation results
of ten different biopsies in which the first (Fig. 17 (a))
and second (Fig. 17 (b)) configurations correspond
respectively to the first (represented by the colored blue
histogram) and second (represented by the colored red
histogram) neuron outputs of the output layer. These
biopsies correspond to the benign class (01 in binary).
On the other hand, Fig. 18 illustrates the DC response
results of ten other different biopsies, in which the first
(represented by the colored red histogram, Fig. 18 (a))
and second (represented by the colored blue histogram,
Fig. 18 (a)) configurations correspond respectively to the (c) (d)
first and second neuron in the output layer. These Fig. 19. AANN DC response for ten close benign and ten close
malignant neuron outputs: (a) First neuron output for ten close benign
biopsies reveal them to be malignant (10 in binary, outputs, (b) Second neuron output for ten close benign outputs, (c) First
corresponding respectively to a 1 for the first neuron and neuron output for ten close malignant outputs, and (d) Second neuron
to a 0 for the second neuron of the output layer). output for ten close malignant outputs.

436
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

TABLE IV: PROCESS CORNER ANALYSIS RESULTS

Case tt ss snfp fnsp ff
Logic 1 1.645 1.523 1.438 1.469 1.417
Logic 0 363.7 375.9 389.0 369.7 399.1

The process corner analysis shown in Table IV verifies

that the output logic values are similar to the results
obtained by using the typical (tt) process corner of
transistors. Moreover, it is evident that the logic 1 limit is
lower and the logic 0 limit is higher for the ff process
corner. These minor changes can be effectively adjusted
by retraining the weights using the backpropagation
algorithm. Therefore, a new weight set that aligns better
with the observed results of the AANN to the design
specifications can be obtained. Thus, it can be inferred Fig. 21. Output prediction curves, with target in red and output in blue.
that AANNs can withstand manufacturing errors and
quickly adapt to new situations.
The power supply rail of each analog circuit and
computing hardware used for the implementation of this
AANN is ±900 mV. Given the analog nature of this
neural network, the circuits require specific voltage levels,
which include positive and negative supply voltages, to
ensure proper biasing and operation of analog circuits,
leading to accurate analog computation. The power
consumption of the analog neural network is shown in
Fig. 20. It is known that the total power consumption of a
circuit is obtained by calculating the area under the graph.
Thus, the total power consumption of this AANN is
31.95 µW. The total elapsed time for the simulation of a (a)
sample in Cadence virtuoso is revealed from its
“psf/spec.” file to be 28.9595 s with a peak resident
memory used of 68.7 Mbytes, yielding a total energy of
0.93 mJ. This very low power consumption is crucial for
ensuring energy efficiency and minimizing heat
dissipation. Hence, it could potentially contribute to a
cost-effective, sustainable, and widely accessible
diagnostic solution for breast cancer.
Fig. 21 illustrates the two outputs neuron prediction
curve for the initial 50 biopsies samples of the analog
implementation. The target output is elegantly depicted
by the bold red curve, while the analog output gracefully
follows the blue curve. These curves encapsulate the
intricate dance of computation and representation within (b)
the neural substrate. This allows efficient and powerful Fig. 22. AANN circular propeller: (a) Analog circular propeller (in blue)
data representation using less power. and (b) Target circular propeller (in red).

Furthermore, Fig. 22 depicts the circular propeller of

the AANN breast cancer classifier. Fig. 22(a) and Fig.
22(b) represent respectively the analog (in blue) and
target (in red) circular propellers. The uniform surface
texture and consistent behavior across various regions
suggest strong generalization capabilities and a well-
fitted model. Implying the model effectively captures
underlying data patterns and provides stable, reliable, and
robust predictions.
The results of this section are used to conclude on the
proposed analog chip, which classifies breast cancer as
either malignant or benign, offering promising
advantages such as energy efficiency, low power
consumption, real-time diagnosis, and potential hardware
Fig. 20. Power consumption of the AANN. optimizations. However, this work equally faces

437
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

difficulties related to precision due to the complexity of measure of the difference between predicted (obtained)
the network and the need for interdisciplinary values and actual values (targets).
collaboration. The next section focuses on the discussion The AANN is trained on the Wisconsin dataset that
of the simulation results. includes women of all ages, and races, as well as those
with various genetic risk factors. It classifies both early-
VI. DISCUSSION stage and late-stage tumors, considering clinical imaging
artifacts such as motion artifacts, superimposed breast
A. Interpretation of the Results tissue and dense breast tissue.
Several different measures are commonly used to
assess the performance measures of a proposed method. B. Validation of the Results
These measures are calculated with the aid of True By firstly subjecting the AANN to Monte Carlo
Positive (TP), False Positive (FP), True Negative (TN), analysis for evaluating process variations by simulating
and False Negative (FN) samples. The extracted 1000 iterations with normally distributed transistor
confusion matrix used to detect cancerous and non- parameters (10% standard deviation), it facilitated the
cancerous breast tumors is clearly shown in Table V. quantification of uncertainties and risks influencing the
network's output predictions, enabling the identification
TABLE V: AANN NETWORK CONFUSION MATRIX
of critical parameters affecting classification. This
Approach AANN classification validation effort not only underscored the model's
Output class non-cancerous cancerous robustness and responsiveness to input fluctuations but
Benign 444 (TN) 14 (FN)
Malignant 8 (FP) 233 (TP) also provided crucial insights for performance
optimization, thereby bolstering network’s confidence
accuracy and applicability in breast cancer diagnosis.
The eight missed benign tumor diagnoses can lead to
Secondly, subjecting the AANN to process corner
unnecessary patient anxiety, increased healthcare costs
variations for assessing extreme conditions by simulating
due to follow-up tests, and potential overdiagnosis. In
at five process corners (tt, ff, ss, snfp, and fnsp), three
contrast, the fourteen missed malignant tumor diagnoses
pose a risk that could delay treatment, worsen patient temperatures (40°C, 25°C, 125°C), and ±10% supply
outcomes, and potentially lead to suboptimal patient care. rail (± 900 mV), facilitated a thorough model's
Classification performance measures include performance evaluation under diverse scenarios. This
sensitivity (SE), specificity (SP), accuracy (AC), and effort does not only underscored network's resilience and
Mathew’s correlation coefficient (MCC). adaptability in environmental challenges but also
provided pivotal guidance for further optimizing the
TP model's performance, therefore instilling confidence in its
SE  applicability for reliable diagnostic in variable
TP  FN
environmental contexts.
TN Thirdly, varying transistor parameters and supply
SP 
TN  FP voltage within a ±10% range revealed that the AANN is
relatively insensitive to these variances. However, the
TP  TN AANN performance is slightly sensitive to temperature
AC 
TP  TN  FN  FP variations, with a 10% temperature increase leading to a
less than 1% decrease in accuracy due to increased noise
 TP  TN    FP  FN  and leakage currents. Mitigation strategies include using
MCC 
 TP  FP  TP  FN  TN  FP  TN  FN  temperature-compensated circuits, or training with data
covering a wide temperature range to enhance robustness.
Furthermore, Table VII compares this work with
TABLE VI: AANN NETWORK PERFORMANCE MEASURES
existing studies. Breast cancer classification is mostly
Approach Number of cases SE SP AC MCC done by ANNs implemented on graphic processing units
AANN 699 0.9433 0.9823 0.9685 0.9309 (GPU), which easily consume huge amounts of power.
TABLE VII: COMPARISON WITH THE STATE OF THE ART
The performance measures shown in Table VI reveal a
96.85% accuracy obtained from Monte Carlo simulation Author
AC
Application Method
Design Supply
Power Energy
results (that is, 22 incorrect logic evaluations of the breast (%) layout rail
cancer classifier out of the 699 Wisconsin dataset). Table M. Kanojia Breast
96.77 GPU -- -- -- --
[52] Cancer
VI equally reveals that the classifier has the advantage of Hua et al. Breast CMOS ±2.500 46.08
obtaining higher speed and density due to a higher 96.43 5-3-1 --
[14] Cancer IC V mW
specificity performance measure. In contrast, it turns to Sanjeev T. Breast CMOS ±1.100 50.00 1.60
96.59 9-5-1
have slight stability challenges with temperature et al. [15] Cancer VLSI V µW mJ
variations due to a sensitivity performance measure close This
96.85
Breast CMOS
9-10-2
±900 31.95 0.93
to, but less than 0.95. Additionally, the AANN has a very work Cancer VLSI mV µW mJ
high network performance, as highlighted by Mathew’s
correlation coefficient of 0.9309, which represents the Unlike studies that focus on digital circuits, the
proposed implemented CMOS circuit classifiers can be

438
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

used to achieve similar tasks. This innovative approach revolutionizing breast cancer screening, enabling faster
offers the potential for improved energy efficiency, faster diagnosis for timely treatment and improved patient
computations, a lesser complexity network, and precise outcomes, reducing healthcare costs, reducing disparities
analog pattern recognition, marking a notable departure in breast cancer care, prioritizing investments in AANN
from digital-centric methodologies. This work bridges the research and development, and establishing policies that
gap between digital-centric methods and promising support the integration of AANNs into clinical practice.
AANNs, leading to improvements in medical diagnostics However, implementing a fully integrated on-chip
tasks, which could have a profound impact on healthcare. learning AANN based on MLP architecture with
Moreover, this study addresses limitations identified in backpropagation can improve the reliability and
previous studies, such as unconventional architecture and efficiency of this work for breast cancer diagnosis.
low noise immunity. It captures complex non-linear
relationships between input and output variables, CONFLICT OF INTEREST
enabling accurate modeling of complex systems. The authors declare no conflicts of interest.
The advantage of this work lies in its pioneering
utilization of analog computing for medical diagnostics. AUTHOR CONTRIBUTIONS
This innovative approach offers a distinct edge in terms
of energy efficiency, real-time processing, and potential The primary investigators and lead authors of the study
for nuanced pattern recognition when compared to the are Djimeli-Tsajio Alain B. and Koagne Longpa T. Silas.
digital-centric methods employed in related studies. By Djimeli-Tsajio Alain B. was responsible for the software
leveraging analog computing principles, this work holds specifications of the AANN, while Geh Wilson Ejuh
promise of enhancing computational speed, potentially played a key role in the AANN’s mathematical modeling.
revolutionizing the medical diagnostic landscape. The Both served in a supervisory capacity. Additionally,
unique capabilities of AANN, including its energy expertise in the training process of the ANN network was
efficiency and potential for real-time diagnosis, mark a provided by Lienou T. Jean-Pierre and Noulamo Thierry.
paradigm shift in the field, positioning this research at the The implementation and testing of the AANN for breast
forefront of cutting-edge advancements in medical cancer diagnosis were carried out by Koagne Longpa T.
diagnostics. Silas. All authors had approved the final version.
Finally, this AANN is a small compact portable
energy-autonomous IC chip of a few dimensions DATA AVAILABILITY STATEMENT
associated with a digital-to-analog converter that can be Data supporting the findings of this study are available
used in laboratories to record breast cancer attributes and from the corresponding author upon reasonable request.
provide appropriate diagnostics. It serves as a diagnostic
decision support system tool in the hospital which will REFERENCES
help the clinicians and physicians in faster decision-
[1] R. Qamar and B. A. Zardari, “Artificial neural networks: An
making and by so doing, will ultimately reduce the risk of overview,” Mesopotamian Journal of Computer Science, pp. 130–
false diagnosis. Thus, hospitals will have to buy the 139, 2023. doi: 10.58496/MJCSC/2023/015
AANN for utilization by physicians, thereby empowering [2] M. Anguita, F. J. Pelayo, A. Prieto, and J. Ortega, “Analog CMOS
physicians with a medical diagnostic cutting-edge tool. implementation of a discrete time CNN with programmable
cloning templates,” IEEE Trans. on Circuits and Systems II:
Analog and Digital Signal Processing, vol. 40, no. 3, pp. 215–218,
VII. CONCLUSION 1993.
[3] H. Chible, M. E. H. Jouni, and B. Hassan-Beirut, “Analog
The design and implementation of an AANN for breast multiplier for feed forward neural network signal processing,” in
cancer cell classification as benign or malignant is Proc. the 7Th Int. Conf. on Signal Processing, Pattern
presented in this paper. The MLP architecture with a back Recognition and Applications, 2010. doi:10.2316/P.2010.678-053
[4] M. Fernando, A. Ana, and M. Alexandre, “Artificial neural
propagation algorithm was retained using MATLAB, networks: A review of commercial hardware,” Engineering
featuring nine inputs, ten neurons in the hidden layer and Applications of Artificial Intelligence, vol. 17, pp. 945–952, 2004.
two neurons in the output layer. Our main contribution doi: 10.1016/j.engappai.2004.08.011
was the development of novel analog circuits for AANN [5] M. Rizwan and Y. R. Alharbi, “Artificial intelligence-based
approach for short term load forecasting for selected feeders at
implementation. Verification and validation occurred in Madina Saudi Arabia,” International Journal of Electrical and
Cadence Virtuoso through Monte Carlo analysis and Electronic Engineering & Telecommunications, vol. 10, no. 5, pp.
process corner variation analysis. The Spice simulation 300–306, 2021.
results aligned with work specifications, showing the [6] A. Frank and A. Asuncion, UCI Machine Learning Repository,
School of Information and Computer Science, University of
final circuit operating at ± 900 mV power supply rail, California, 2010.
consuming 31.95 µW. The “psf/spec.” simulation time [7] H. Jouni, “Cellules analogiques CMOS pour réseaux de neurones.
was 28.9595 seconds per sample, with a peak resident Application à la classification des cellules cancéreuses dans le
memory usage of 68.7 Mbytes, yielding a total energy of sein,” dissertation, COMUE Université Côte d’Azur, Lebanese
International University, 2018.
0.93 mJ. The AANN achieved a 96.85% accuracy and [8] H. Jouni, A. Harb, G. Jacquemod, and Y. Leduc, “Programmable
0.9309 MCC, promising for breast cancer classification. signal generator for neural network application,” in Proc. 2017
This work holds the impact of providing a solid 29th Int. Conf. on Microelectronics, 2017. doi:
background for future studies, opens up new possibilities 10.1109/ICM.2017.8268881
[9] H. Jouni, A. Harb, G. Jacquemod, and Y. Leduc, “Creation of real
for optimizing computational resources and

439
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

blocks for neural network using Simulink,” in Proc. 2018 Int. based models with conventional regression models for predicting
Conf. on Computer and Applications, 2018. doi: spray drift,” Environmental Science and Pollution Research, vol.
10.1109/COMAPP.2018.8460327 30, no. 8, pp. 21927–21944, 2023.
[10] H. Jouni, A. Harb, G. Jacquemod et al., “Design and specification [29] J. Sedlakova, A. L. Westermair, N. Biller-Andorno, C. A. Meier,
of analog artificial neural network,” SN Applied Sciences, vol. 1, and M. Trachsel, “Comparison of analog and digital patient
no. 11, pp. 1–15, 2019. doi: 10.1007/s42452-019- 1243-4 decision aids for the treatment of depression: A scoping review,”
[11] H. Jouni, A. Harb, G. Jacquemod, and Y. Leduc, “Design and Frontiers in Digital Health, vol. 5, #1208889, 2023.
specification of analog artificial neural network,” SN Applied [30] D. K. McNeill, C. R. Schneider, and H. C. Card, “Analog CMOS
Sciences, vol. 1, no. 11, pp. 1–15, 2019. neural networks based on Gilbert multipliers with in-circuit
[12] Z. Jia and B. Kim, “Online trained neural network-PI speed learning,” in Proc. 36th Midwest Symp. on Circuits and System,
controller for DTC based IPMSM drives,” International Journal 1993, pp. 1271–1274.
of Electrical and Electronic Engineering & Telecommunications, [31] S. L. Pinjare, “Design and analog VLSI implementation of neural
vol. 7, no. 3, pp. 108–113, 2018. network architecture for signal processing,” European Journal of
[13] A. Jayaraj, I. Banerjee, and A. Sanyal, “Common-source amplifier Scientific Research, vol. 27, no. 2, pp 199–216, 2009.
based analog artificial neural network classifier,” in Proc. 2019 Int. [32] B. D. Yammenavar, V. R. Gurunaik, R. N. Bevinagidad, and V. U.
Symp. on Circuits and Systems, pp. 1–5, 2019. doi: Gandage, “Design and analog VLSI implementation of artificial
10.1109/ISCAS.2019.8702426 neural network,” International Journal of Artificial Intelligence &
[14] M. S. Alejandro, C. A. Hernandez-Gracidas, L. A. Morales- Applications (IJAIA), vol. 2, no. 3, pp. 96–109, 2011.
Rosales, I. Algredo-Badillo, M. Amador Garcıa, and J. A. Orozco [33] S. C. Liu, J. P. Strachan, and A. Basu, “Prospects for analog
Torres, “CMOS implementation of ANNs based on analog circuits in deep networks in analog circuits for machine learning,
optimization of n-dimensional objective functions,” Sensors, vol. current/voltage/temperature sensors, and high-speed
21, no. 21, p. 7071, 2021. communication: Advances in analog circuit design,” Cham., pp.
[15] T. S. Chandrasekaran, R. Hua, I. Banerjee, and A. Sanyal, “A 49–61, 2021. doi: https://doi.org/10.1007/978-3-030-91741-8_4
fully-integrated analog machine learning classifier for breast [34] A. F. Murray, D. Del Corso, and L. Tarassenko, “Pulse-stream
cancer classification,” Electronics, vol. 9, no. 3, p. 515, 2020. VLSI neural networks mixing analog and digital techniques,”
[16] A. Arabi, N. Bourouba, A. Belaout, and M. Ayad, “An accurate IEEE Trans. on Neural Networks, vol. 2, no. 2, pp. 193–204, 1991.
classifier based on adaptive neuro-fuzzy and features selection [35] W. P. Chen, T. Y. Wang, H. L. Xu, and X. W. Liu, “A four-
techniques for fault classification in analog circuits,” Integration, quadrant analog multiplier based on CMOS source coupled pair,”
vol. 64, 2019. doi: https://doi.org/10.1016/j.vlsi.2018.08.001 Key Engineering Materials, vol. 483, pp. 487–491, 2011. doi:
[17] B. Gao, Y. Zhou, Q. Zhang, S. Zhang, P. Yao, Y. Xi, Q. Liu, M. 10.4028/www.scientific.net/KEM.483.487
Zhao, W. Zhang, Z. Liu, X. Li, J. Tang, H. Qian, and H. Wu, [36] S. Panda, S. Srimani, B. Maji, and A. Mukhopadhysay, “High
“Memristor-based analogue computing for brain-inspired sound performance CMOS four quadrant analog multiplier in 45 nm
localization with in situ training,” Nature Communications, vol. technology,” International Journal of Application or Innovation in
13, no. 1, 2022. doi: https://doi.org/10.1038/s41467-022-29712-8 Engineering & Management, vol. 2, pp. 1–7, 2013. doi:
[18] W. Xu, J. Wang, and X. Yan, “Advances in memristor-based 10.1109/JSSC.1986.1052546
neural networks,” Frontiers in Nanotechnology, vol. 3, 645995, [37] S. L. Pinjare and P. C. P. Raj, “Design and analog VLSI
2021. doi: 10.3389/fnano.2021.645995 implementation of neural network architecture for signal
[19] N. Tran-Thi-Kim, T. Pham-Viet, I. Koo, V. Mariano, and T. Do- processing,” European Journal of Scientific Research, vol. 27, no.
Hong, “Enhancing the classification accuracy of rice varieties by 2, pp. 199–216, 2009.
using convolutional neural networks,” International Journal of [38] V. Nandhini and K. Sambath, “VLSI implementation of multiplier
Electrical and Electronic Engineering & Telecommunications, vol. design using reversible logic gate,” Analog Integrated Circuits
12, no. 2, pp. 150–160, 2023. and Signal Processing, vol. 115, no. 1, pp. 93–100, 2023.
[20] M. Puttagunta and S. Ravi, “Medical image analysis based on [39] G. Han and E. Sanchez-Sinencio, “CMOS transconductance
deep learning approach,” Multimedia Tools and Applications, vol. multipliers: A tutorial,” IEEE Trans. on Circuits and Systems II:
80, pp. 24365–24398, 2021. Analog and Digital Signal Processing, vol. 45, no. 12, pp. 1550–
[21] H. Jouni, M. Issa, A. Harb, G. Jacquemod, and Y. Leduc, “Neural 1563, 1998.
network architecture for breast cancer detection and [40] R. Wojtyna and T. Talaska, “Transresistance CMOS neuron for
classification,” in Proc. 2016 Int. Multidiscip. Conf. on adaptive neural networks implemented in hardware,” Bulletin of
Engineering Technology, 2016, pp. 37–41. the Polish Academy of Sciences: Technical Sciences, vol. 54, no. 4,
[22] F. B. Gencer, X. Xhafa, B. B. İnam, and M. B. Yelten, “Design pp. 443–448, 2006.
and validation of an artificial neural network based on analog [41] Z. You and C. Lu, “A heuristic fault diagnosis approach for
circuits,” Analog Integrated Circuits and Signal Processing, vol. electro-hydraulic control system based on hybrid particle swarm
106, pp. 475–483, Marr. 2021. optimization and Levenberg–Marquardt algorithm,” Journal of
[23] K. Freeman, J. Geppert, C. Stinton, D. Todkill, S. Johnson, A. Ambient Intelligence and Humanized Computing, pp. 1–10, 2018.
Clarke, and S. Taylor-Phillips, “Use of artificial intelligence for doi: https://doi.org/ 10.1007/s12652-018-0962-5
image analysis in breast cancer screening programmes: Systematic [42] P. Zhou and J. Austin, “Learning criteria for training neural
review of test accuracy,” BMJ, vol. 374, 2021. doi: network classifiers,” Neural computing and applications, vol. 7,
https://doi.org/10.1136/bmj.n1872 pp. 334–342, 1998. doi: https://doi.org/10.1007/BF01428124
[24] R. Mina, C. Jabbour, and G. E. Sakr, “A review of machine [43] J. Spiegel, P. Mueller, D. Blackman et al., “An analog neural
learning techniques in analog integrated circuit design computer with modular architecture for real-time dynamic
automation,” Electronics, vol. 11, no. 3, p. 435, 2022. doi: computations,” IEEE Journal of Solid-State Circuits, vol. 27, no. 1,
https://doi.org/10.3390/electronics11030435 pp. 82–92, 1992
[25] E. H. Houssein, M. M. Emam, and A. A. Ali, “An optimized deep [44] MathWorks MATLAB and statistics toolbox release 2016a, The
learning architecture for breast cancer diagnosis based on MathWorks Inc, Natick, Massachusetts, United States, 2016.
improved marine predators’ algorithm,” Neural Computing and [45] C. Mead, “Neuromorphic electronic systems,” in Proc. the IEEE,
Applications, vol. 34, no. 20, pp. 18015–18033, 2022. vol. 78, no. 10, pp. 1629–1636, 1990.
[26] N. Ye, L. Cao, L. Yang, Z. Zhang, Z. Fang, Q. Gu, and G. Yang, [46] G. I. Salama, M. Abdelhalim, and M. A. Zeid, “Breast cancer
“Improving the robustness of analog deep neural networks through diagnosis on three different datasets using multi-classifiers,”
a bayes-optimized noise injection approach,” Communications International Journal of Computer and Information Technology,
Engineering, vol. 2, no. 1, #25, 2023. vol. 1, no. 1, pp. 36–43, 2012.
[27] I. B. Ahmed, W. Ouarda, and C. B. Amar, “Hybrid UNET model [47] N. H. Serajee, S. Mannan, R. Abdulghafor, S. Wani, A. A.
segmentation for an early breast cancer detection using ultrasound Ibrahim, and A. Olowolayemo, “Breast cancer prediction using
images,” in Proc. Int. Conf. on Computational Collective machine learning,” International Journal on Perceptive and
Intelligence, 2022, pp. 464–476. Cognitive Computing, vol. 8, no. 1, pp. 24–28, 2022.
[28] G. Moges, K. McDonnell, M. A. Delele, A. N. Ali, and S. W. [48] M. F. Aslan, “A hybrid end-to-end learning approach for breast
Fanta, “Development and comparative analysis of ANN and SVR- cancer diagnosis: Convolutional recurrent network,” Computers

440
 International Journal of Electrical and Electronic Engineering & Telecommunications Vol. 13, No. 6, 2024

and Electrical Engineering, vol. 105, 108562, 2023. doi: publications in the area of method engineering, multi-agent systems,
https://doi.org/10.1016/j.compeleceng.2022.108562 neural networks, and computer vision. The recent research interests
[49] H. Hakkoum, A. Idri, and I. Abnane, “Assessing and comparing concern both some open questions concerning cyber resilience and
interpretability techniques for artificial neural networks breast Machine Learning. He is a member of the International Association of
cancer classification,” Computer Methods in Biomechanics and Engineers. He received the first Cameroon National Prize of Innovation
Biomedical Engineering: Imaging & Visualization, vol. 9, no. 6, in Embedded Systems and is currently a peer reviewer in a few
pp. 587–599, 2021. International Journal of Computer Science.
[50] I. AlShourbaji, P. Kachare, W. Zogaan, L. J. Muhammad, and L.
Abualigah, “Learning features using an optimized artificial neural Noulamo Thierry earned his master’s degree,
network for breast cancer diagnosis,” SN Computer Science, vol. 3, Diploma of Advanced Studies (DEA), and
no. 3, 2022. doi: https://doi.org/10.1007/s42979-022-01129-6 Ph.D. in software engineering science from the
[51] A. S. Orgenci, G. Dundar, and S. Balkur, “Fault-tolerant training University of Yaoundé I in 2000, 2001, and
of neural networks in the presence of MOS transistor 2010, respectively. He is currently working as
mismatches,” IEEE Trans. on Circuits and Systems II: Analog and a lecturer in the Department of Computer
Digital Signal Processing, vol. 48, no. 3, pp. 272–281, 2001 Engineering at Fotso Victor University
[52] M. Kanojia, “Malignancy detection in breast histo-images using Institute of Technology (IUT-FV) of Bandjoun,
multi-layer perceptron,” in Proc. Int. Conf. on Soft Computing and since 2004. He has been a member of EANG
Pattern Recognition, pp. 553–562, 2021. since 2009, Life member of UR.A.I.A. since 2008. He has published
more than 11 research papers in reputed international journals and
Copyright © 2024 by the authors. This is an open access article conferences. His main research work focuses on Process Automation
distributed under the Creative Commons Attribution License (CC BY- using MDE and Muti-Agent approach. He has 18 years of teaching
NC-ND 4.0), which permits use, distribution, and reproduction in any experience and 11 years of research experience.
medium, provided that the article is properly cited, the use is non-
commercial and no modifications or adaptations are made. Geh Wilson Ejuh is a distinguished academic
and researcher in the field of electrical and
Djimeli Tsajio Alain Bernard received a B.S. electronic engineering. He embarked on his
(2001) and M.S. with thesis (2004) from the academic journey at the University of Yaounde
Faculty of Science of the University of I, where he earned a BSc in physics-chemistry
Yaoundé I and the Ph.D. (2016) from the in 2003, followed by an MSc in physics with a
Faculty of Science of the University of focus on materials science in 2006. After
Dschang, all in Cameroon in the field of working as a part-time lecturer in the Physics
physics option Electronics. Since 2006, he has Department of ENS Bambili from 2007, his
joined Fotso Victor University Institute of career as an educator began at Gombe State
Technology of the University of Dschang as a University in Nigeria, where he served as an
lecturer in the Department of Assistant Lecturer from November 2008. He then joined the teaching
Telecommunication and Network Engineering. He is a member of staff of IUT-FV Bandjoun, University of Dschang as an Assistant
UR.A.I.A. of the present university where he is carrying out research in Lecturer in 2011, and obtained his PhD in 2013 in physics option
the field of Artificial Intelligence for biomedical process. He is the materials sciences from the Faculty of Science of the University of
author of more than 10 research papers. Yaoundé 1, Cameroon. In 2014, he was promoted to the rank of
Lecturer, in 2019 to the rank of Associate Professor and in 2023, he was
Koagne Longpa T. Silas earned his promoted to the rank of Professor of Physics. He was posted as Head of
Technical School Teacher’s Training Diploma Services for Teaching and Academic Activities in NAHPI in 2017 and
I (DIPET I) in Electrical and Electronics in 2021 as Head of Services for continuous training in NAHPI,
Engineering, option Electronics at the University of Bamenda. Currently, he contributes his expertise at the
University of Bamenda in 2018. His University of Dschang, IUT-FV Bandjoun, Cameroon, within the
unwavering commitment to academic General and Scientific Studies department, and at the National Higher
excellence led him to obtain a Technical Polytechnic Institute of the University of Bamenda, in the Department
School Teacher’s Training Diploma II of Electrical and Electronic Engineering. His research is at the forefront
(DIPET II) in Electrical and Electronics of exploring the electronic, optoelectronic, photonic, and
Engineering, option Electronics at the University of Bamenda in 2020. thermodynamic properties of organic and pharmaceutical molecules.
His deep-rooted passion for research in electrical and electronics His work on the doping of fullerene and other organic materials is
engineering motivated him to further his studies at the University of particularly notable. He established himself as an expert in Material
Dschang where he obtained a bachelor’s degree in sciences and Characterization, Optoelectronics, Photovoltaics, and Nanotechnology
technology (BSc) in physics, option Electrical, Electronics, and with a prolific output of 80 publications, which have garnered 16,782
Automation (EEA) in 2021 and followed by a master degree of sciences reads and 741 citations. His contributions to the scientific community
and technology (MSc) in physics option Electrical, Electronics and are substantial, standing at the head of the Quantum Dynamics Research
Automation (EEA) in 2022. He currently stands at the forefront of Group (QDRG) at ICTP-EAIFR, and being a member of reviewer to
technology and scientific research as a UR.A.I.A. member and Ph.D. many Journals. His dedication to advancing the field of materials
Student in Physics specializing in Electronics at the University of science and engineering is evident in his impressive academic and
Dschang. research accomplishments.

Lienou T. Jean-Pierre, born in Kekem in

1966. After studying Russian language in 1988
at Kharkiv Polytechnic Institute, He obtained
his MSc in systems and networks at Kyiv
Polytechnic Institute of the National Technical
University of Ukraine in 1994. After working
sometimes in private companies mainly as
Biomedical Engineer or System Engineer, He
joined the teaching staff of the University of
Dschang as Assistant Lecturer in 2002 and
obtained his Ph.D. in 2015 at Yaoundé National Advanced School of
Engineers of the University of Yaoundé I, Cameroon. He is now an
associate professor of Computer Science and was posted as head of the
Department of Computer Engineering in the College of Technology at
the University of Bamenda in 2017. Author and co-author of

441