An enhanced UNet variant for

Effective Lung Cancer Detection

Lerina Aversano Mario Luca Bernardi
Dept. of Engineering Dept. of Engineering
University of Sannio University of Sannio
Benevento, Italy Benevento, Italy
[email protected] [email protected]

Marta Cimitile Martina Iammarino Chiara Verdone

Dept. of Law and Economics Dept. of Engineering Dept. of Engineering
UnitelmaSapienza University University of Sannio University of Sannio
2022 International Joint Conference on Neural Networks (IJCNN) | 978-1-7281-8671-9/22/$31.00 ©2022 IEEE | DOI: 10.1109/IJCNN55064.2022.9892757

Rome, Italy Benevento, Italy Benevento, Italy

[email protected] [email protected] [email protected]

Abstract—Lung cancer is one of the diseases with the early stage. There are several tests that can determine
highest mortality rate and early detection is key to making the health of the lungs, but the test considered most
the disease as treatable as possible. The most common effective at an early stage is known as low-dose CT
and useful method for recognizing pulmonary nodules is
computed tomography, which allows them to be located
(computed tomography). The radiologist analyzes the
and monitored. The disadvantage of this technique is that images produced to look for abnormal areas in the lungs
the scans have to be interpreted by doctors, who could that could indicate cancer. A spot visible in the image
make mistakes. The use of pulmonary CAD is therefore is known as a "pulmonary nodule" and needs to be
becoming increasingly widespread, a system capable of au- monitored. The problem is represented by the criticality
tomatically analyzing CT images and providing information of the interpretative analysis by the doctor who finds
on possible suspicious regions found in the images. These
systems, by offering radiologists a list of already marked himself evaluating a high number of images with the
regions of interest to view with particular attention, increase risk of making mistakes and causing incorrect diagnoses.
the efficiency of detection of small nodules and reduce This explains the introduction of some computerized
reporting times by physicians. This study aims to accurately image analysis systems, called CAD (Computer-Aided
detect the location of pulmonary nodules through a Deep Detection), which provide the localization and automatic
Learning approach with the use of computed tomography
identification of the pulmonary nodule. This method
scans. In particular, it proposes the use of a new variant
of the UNet architecture, called GUNet3++, which has been can significantly reduce errors, reducing time [12], and
compared with the other types of this network. To validate identifying half of the cases lost to humans [31]. In recent
the approach, the public LIDC-IDRI dataset was used, which years, several CAD systems have been developed for
collects pulmonary CT images of about a thousand patients the detection and classification of pulmonary nodules in
with different types of cancer. The results obtained are very CT [13], [30] imaging systems, some based on machine
promising, showing a performance improvement compared
to other UNet networks.
learning [32] and others on Deep Learning (DL) ap-
Index Terms—Deep Learning, U-Net, CT scan Images, Lung proaches [22]. Among the latter, which show better per-
Cancer formance than traditional lung nodule detection systems,
an example is the use of CT scan images for pulmonary
I. I NTRODUCTION nodule detection [1]. In particular, in [4] lobe-driven CT
image clustering is used for classification.
Lung cancer is a very serious disease that affects a
vital organ for the human being, and unfortunately, it Like the previous ones, this study also proposes a DL-
represents the first cause of death from cancer in indus- based approach to automatically and accurately detect
trialized countries [18]. Its incidence is constantly and lung cancer by performing the semantic segmentation
continuously increasing [19], therefore early diagnosis of the CT scans. More specifically, we use a generalized
can make a difference by allowing prompt treatment. UNet-based architecture, called GUNet3++, that includes
In this regard, it is of fundamental importance to sub- multiscale and dense skip connections to build a more
ject high-risk individuals to regular screenings [8], [20], accurate model of the detected lung nodules.
thanks to which it is possible to detect cancer at a very The document is structured as follows: in Section II

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on February 11,2025 at 04:23:51 UTC from IEEE Xplore. Restrictions apply.
978-1-7281-8671-9/22/$31.00 ©2022 IEEE
the most relevant related works are reported, Section III The proposed approach aims to overrun these limitations
concerns the pre-trained neural networks adopted, some by improving the semantic segmentation performance
fundamental concepts are in fact explained, and Section with a more effective feature extraction at both intra and
IV describes the approach used in detail. The description inter-resolutions levels.
of the experiment is in Section V, and Section VI shows
the results obtained. Finally, Section VII and VIII report III. B ACKGROUND
threats to validity and conclusions, respectively. This study aims to perform the semantic segmentation
of CT scans to identify lung cancer lesions.
II. R ELATED W ORK Semantic segmentation is the process of partitioning
In recent years, Machine Learning and Deep Learning an image into meaningful regions where each pixel in
(DL) techniques have been increasingly adopted to deal a region is similar to the others in the same region in
with mining and classification problems in the medical some properties or characteristics. The attention to the
field [2], [3], [10]. categorization of the pixels is very relevant as a support
In their work [33], Zhang et al. present a very detailed in the medical field as it represents a diagnostic tool, for
systematic review of existing techniques that aim to di- example, a nodule is identified by pixels belonging to
agnose nodules using CT images. This review shows that the same class [21].
many approaches use 2D CNN, DBN, and MTANN [27], The proposed approach is based on variants of the
[28], although these are superficial and have daunting UNet network, called GUNET3++, therefore we briefly
performance. illustrate the main variants of the UNet network that led
Starting from this assumption and given that the DL to the creation of the network used.
has a better classification capacity than the classical lung
cancer detection techniques [9], the next step in the A. UNET
development of the DL techniques is the optimization UNet is a fully convolutional [24] neural network,
of network performance [14], [25]. whose architecture has been developed to be able to
In the work [17], the authors combine DL techniques work with fewer training images and produce more pre-
with additional information, such as clinical factors, in cise segmentation, even in the 3D imaging segmentation
order to reduce possible classification errors; the ultimate field [5].
goal is early diagnosis. The idea behind UNet is to add successive layers
To achieve the same goal, Gao et al. [11] instead use to a standard contractual network where upsampling
the deep neural networks CNN, DNN, and SAE, on the operations will take place, increasing the resolution of
LIDC-IDRI database [16], obtaining an accuracy equal to the output. This way a subsequent convolutional layer
about 85%. takes information on how to assemble the next output.
The study proposed in [35] aims at the classification It has a U-shaped architecture and consists of a con-
of nodules from CT scans, but it uses the DeepLung tracting path and an expansive path. The first is a typical
approach and is based on two different phases. The first convolutional network in which each convolution is
phase concerns the detection of candidate nodules based followed by a rectified linear unit (ReLU) and a max-
on a 3D Faster R-CNN and a second phase in which pooling operation. Therefore, the information on the
the deep characteristics are extracted and classification characteristics is increased while the spatial information
is carried out through a DPN network. is reduced. The second path has the task of combining
In the work [1], the authors aim to specialize classifiers spatial information and characteristics through convo-
on a specific lobe of the lung using an overall approach lutions that use high-resolution characteristics deriving
based on image grouping. Different pre-trained convo- from the contraction path.
lutional neural networks were used, the best classifier The network uses only the valid part of each con-
was selected for each. The resulting accuracy in the volution with no fully connected layer. Edge pixels are
case of the classification of the single image is equal predicted based on the input image. This architecture is
to 96%, while in the classification of the patient it is suitable for working with large images and allows you
equal to 94%. In case a small number of images is to improve image resolution by overcoming the limit of
available, it is possible to use approaches based on U- GPU memory.
Net [23] which manage to obtain good performances.
Therefore, there are several works based on the seg- B. UNET+ and UNET++ variants
mentation of images [6], [26]. Note that performance Zhou et al. in [34] have introduced three different
may vary based on the size of the receptive field in variants of the UNet network. The first is UNete , made
the convolution kernel. Unfortunately, the use of U-Net up of an overall architecture made up of UNets of dif-
networks has limitations, the fixed receptive field of the ferent depths that partly share the same codifiers while
kernel causes a poor diversity of characteristics, thus each has its decoder. This implies that the following
semantically different characteristic maps are obtained. networks do not supervise the decoders of the previous

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on February 11,2025 at 04:23:51 UTC from IEEE Xplore. Restrictions apply.
 Deep Supervision Deep Supervision

Lfunc Lfunc

320
64

128 320

256 320

320
512

1024

(a) UNet++ variant (b) UNet3+ variant

Fig. 1: Variants of UNet

ones, and the skip connections combine the maps of the vision extended on several scales for the training phase,
characteristics of the decoder only in the same resolu- and modifying the skip connections by combining the
tion scale resulting in too restrictive. The problem of multiscale functionalities. It uses a different approach
skip connections has been solved in the second version, than previous versions as it uses information from all
UNet+, where skip connections become direct between available scales. Each node of the decoder is capable
two adjacent nodes. This also solves the problem of no of acquiring maps of the characteristics of the same
supervision signal between deep and shallow decoders. scale, of a higher and lower scale. In this way, the
The presence of direct skip connections allows training deep and superficial semantic data are combined using
the UNet + network both in a classic way and with in- a feature aggregation mechanism and it is possible to
depth supervision. Finally, UNet++ is constituted starting obtain coarse and fine-grained semantic information on
from UNet+ connected to the decoders through densely all the resolution scales analyzed. The peculiarity of this
skip connections, resulting in a more flexible fusion of architecture is shown in Figure 1-(b), where it is possible
functionality at the decoder nodes. Each node of the to notice that the network, on all available scales, learns
decoder performs two different operations: it horizon- both fine-grained and coarse-grained information. The
tally combines the characteristics at the same resolution decoder, at each level, acquires maps of the same scale
from all the previous nodes and vertically integrates from the encoder, those of the smallest characteristics
the multi-scale functionalities on different resolutions of and those of the largest characteristics of the decoder.
the previous nodes. In this way, maps of the aggregate
characteristics are obtained that generate a more robust IV. T HE GUN ET 3++ APPROACH
training process, less loss of semantic information, and This work aims at the precise and automatic detection
greater accuracy. In Figure 1-(a), we report the architec- of lung nodule position using CT images. The approach
ture scheme in which the decoder nodes are indicated is based on DL, in particular, the GUNet3++ network is
with D, the encoders with E, and those of the pyramidal used, for the CT-scan semantic segmentation.
block with N. The diagram shows how nodes of level The new variant of the UNet network proposed is a
(l) greater than 1 receive l + 1 input. Of these inputs, l compromise between the UNet++ and UNet3+ networks
corresponds to the outputs of the previous levels, while and is called GUNet3++. As already mentioned in the
(l + 1)th corresponds to the skip connection. Also, in the previous section the UNet++ network adds dense skip
figure, the skip connections are shown in blue. connections on each scale, while UNet3+ performs an
aggregation operation of multiscale functionality on each
C. UNET3+ variant decoder. The GUNet3++ architecture was born as a com-
UNet3+ is another variant that derives from UNet, bination of the two networks mentioned and includes
which has the particularity of exploiting the deep super- the strengths of each. As can be seen from Figure 2,

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on February 11,2025 at 04:23:51 UTC from IEEE Xplore. Restrictions apply.
 Deep Supervision N11 , and N21 , all channels are upscaled or downscaled.
Defining i as the down-sampling layer on the
encoders path, j as the convolution layer across the
skip connection, and N + 1 as the number of scales, the
output of the generic node Y i,j is defined as:
  


Hc [y i,k ]j−1
k=0 , U (y
i+1,j−1
) i∈[0,N/2]
j∈[1,N −i−1]





 
Hc [y i,k N −i−1 , C D y i−(k+1),j+k

N −1−j ,

 i∈[0,N −1]
512 k=0 k=0 j=N −i

64 y i,j = 

N −(i+1)
C U y i+(k+1),j−(k+1) ) k=0








448

i>0


128 Hc (D y i−1,j )


j=0

where C(·) is the convolution operation, Hc (·) is the

384 convolution followed by the activation operation, D(·)
256
is the down-sampling operation,   U (·) represents the
up-sampling operation, and · is the concatenation
320
operator.
512 The network was trained using a hybrid loss function
that takes into account both the soft dice coefficient loss
and the pixel-wise cross-entropy loss. Like the other
1024
variants, it is possible to carry out deep supervision on
several scales in order to learn the hierarchical repre-
Fig. 2: The GUNet3++ architecture.
sentations from the maps of the characteristics of all the
resolutions. For this reason, the loss function GLoss is
defined as a weighted sum of an hybrid loss of each
Downscaling 2x 64, 3x3 Conv
decoder:
ND
X
64, 3x3 Conv GLoss = DLoss (G, Y k )
k=1
64, 3x3 Conv
where k is an index across the ND decoders, (G, Y k ) are
64, 3x3 Conv the true labels and the labels evaluated by decoder k for
every pixels in the batch, and DLoss (G, Y k ), defined as
448, 3x3 Conv the sum of pixel-wise cross-entropy and dice-coefficient
Batch Normalization

Upscaling 2x
ReLU contribution, can be fomulated as:
64, 3x3 Conv
NC X
PS  
X L(pn , c) ∗ y(pn , c)
DLoss (G, Y k ) = −1/PS 2∗ + L(pn , c) ∗ log(y(pn , c))
L(pn , c)2 + y(pn , c)2
Upscaling 4x 64, 3x3 Conv c=1 n=1

where, c is one among the NC classes, pn are the PS

Upscaling 6x 64, 3x3 Conv pixels belonging to the current batch, L(pn , c) ∈ G are
the target labels, y(pn , c) ∈ Y k are the predicted labels
by the decoder k.
Fig. 3: Aggregated features map for decoder D31
V. E XPERIMENT D ESCRIPTION
In this section, we describe the adopted CT scans
the architecture is able to propagate information at any dataset and the details on the implementation and vali-
scale from shallow to deeper nodes through the dense dation of the proposed model.
pyramidal block of transducers. In addition, the network
is also able to learn from different regions thanks to the A. Dataset
multiscale skip connections. The dataset adopted is the Lung Image Database Con-
As an example of possible operation of the decoder, sortium image collection (LIDC-IDRI) [16], an interna-
we show the Figure 3, where we want to generate the tional resource accessible from the web. It contains diag-
output signal to the decoder node D31 through upscaling nostic and screening scans of lung cancer with chest CT
and downscaling. A set of six feature maps were used to from 1010 patients. For each image, a double-check was
achieve the output in this case. The resulting feature map carried out. In the first phase, each doctor independently
is 448 filters obtained by convolution with 64 filters of examined and marked the lesions, in the second phase
size 3x3. The figure shows that, with the exception of E30 , the doctors compared their evaluations and expressed

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on February 11,2025 at 04:23:51 UTC from IEEE Xplore. Restrictions apply.
a final opinion. More specifically, a nodule is indicated value. The dropout rate (D0) is the fraction of neurons to
when everyone agrees on the existence of a lesion whose be reset in a given level at each training phase. The deep
diameter exceeds 3 mm. In addition, for each patient im- supervision (DS) indicates precisely whether the internal
age, there is also a metadata file that records the results nodes of the network are also included in the loss or if
of an image annotation process performed by the four only the output has been considered [15]. In the first
expert radiologists. The annotation is aimed to identify case, this parameter has the value yes, otherwise it does
the nodule position in the image. It is described using not. The batch size (BS) indicates the number of samples
the metric Nodule Contour ROI, giving a complete three- processed before the model was updated. Its size must
dimensional contour of the nodule [16], thanks to which be greater than or equal to one and less than or equal to
it is possible to obtain a mask for each representative the number of samples in the training dataset. Finally,
image of the nodules present, and an empty mask for batch normalization manages the normalization of the
images where it is not indicated the presence of nodes. level inputs by re-centering and resizing and is used in
Scans that had low quality and all images belonging to 3 configurations: H, if placed at the beginning of the
patients who did not have enough scans were removed block, T at the end, N if not used.
from the original dataset.
C. Validation
Therefore, three datasets were built, the Oracle Nod-
ules containing the 32,606 images of the LIDC-IDRI To validate the model we have adopted a metric that
database of 500 patients with related metadata, with is usually used to evaluate segmentation: the sørensen-
BMP format and resolution of 512×512 pixel (i); the Dice similarity coefficient, which measures the similarity
CT scan training data set containing the same images between two samples and is based on presence and
as the previous one, of which 28,227 do not contain absence data [7], and is similar to the Intersection over
nodules (ii); and the CT scan test dataset containing Union (IoU) metric, a good metric for measuring the
images of the remaining 500 patients (iii). Notice that overlap between two bounding boxes or masks [29].
the masks obtained by the ROI metrics are necessary Defined G as the correct mask and P as the segmenta-
for the training of the UNET network. However, the CT tion generated by the network, the Dice can be defined
scans and the corresponding masks are used to train the as:
network to perform the semantic segmentation of the 2 ∗ |G ∩ P |
nodules. Dice =
|G| + |P |
B. Experimental setting To validate pixel classification performance and cor-
rectness of the segmentation area we adopted the sen-
The experiment conducted had a dual purpose, to
sitivity (SEN) and the positive predictive value (PPV)
detect the region of interest of the lung lesion in CT
metrics, which are defined as:
scans with the highest possible accuracy rate, through an
approach based on the use of a GuNet3++ network, and |G ∩ P |
to use window correction to evaluate the performance SEN =
|G|
in identifying nodules in the different images.
The experimentation was conducted thanks to the |G ∩ P |
PPV =
use of an Intel Core i9 7920X with 18 cores, 128GB |P |
of RAM, and two 24Gb NVIDIA RTX 3090 GPUs. The VI. R ESULTS AND D ISCUSSION
classifiers used were implemented in Python, through
The table I summarizes the results obtained for
the use of Tensorflow 1 , an open-source software library
GUNET3++ compared to the considered baseline meth-
for machine learning, and Keras 2 , a powerful and easy-
ods (i.e., UNET, UNET++ and UNET3++). The training
to-use open-source Python library for developing and
process uses a hyperparameter optimization step (HPO)
evaluating deep learning models.
on the intervals shown in the II table, so the table shows
The classifiers were trained by optimizing the hy-
the best five hyperparameter configurations. To demon-
perparameters to find the best possible configuration.
strate the effectiveness of the GUNET3++ network, we
More specifically, the hyperparameters tuned are re-
too applied the same hyperparameter optimization to all
ported in Table II, where in the second column there
basic methods to select the best-optimized network of
are the selected ranges adopted. Respectively, first (LR)
each type. Above, in the table, for each type of network,
represents the step size at each iteration as it moves
we report the number of parameters used, in the first
towards a minimum loss function. The optimizer allows
5 columns we report the hyperparameters used, except
identifying, through a series of iterations, those weight
for the UNET network, for which there are four columns
values such that the cost function has the minimum
because in-depth supervision is not used as it is not
1 https://www.tensorflow.org/ applicable. The last three columns show the performance
2 https://keras.io/ of the proposed approach, evaluated using the above

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on February 11,2025 at 04:23:51 UTC from IEEE Xplore. Restrictions apply.
TABLE I: Results comparison of GUNET3++ with baseline methods (best five configuration of hyper-parameters).

Hyperparameters Ranges
Learning rate [0.01, 0.05]
configurations is significantly higher meaning that the
Optimization algorithm SGD, Adam , RmsProp, Nadam training process is less stable. The other two networks
Dropout rate [0.10, 0.25] (UNET++ and UNET3+) provide much better results
Deep supervision Yes, No with several configurations reaching a Dice of ≈ 0.92 for
Batch size { 16, 32, 64, 128, 256}
UNET++ and ≈ 0.93 for UNET3+. As the figure shows,
Batch Normalization Head, Tail, None
GUNET3++ sits in the middle for training stability, hav-
TABLE II: Hyper-parameters Optimization and selected ing a lower number of unsuccessful configurations than
ranges. UNET but higher concerning UNET++ and UNET3+.
This is due to the number of parameters that are sub-
stantially higher in GUNET3++ with respect to other
networks, making the training process more difficult and
less stable. Conversely, GUNET3++ is the best perform-
0.9
ing network in terms of Dice having a high number of
configurations reaching an average Dice equals to ≈ 0.96.
Dice distribution

0.8 This is also noticeable by the qualitative analysis

shown in Fig.5 where the reference annotations agreed
by expert radiologists are compared with ROI extracted
0.7
by predicted masks. As the figure shows, the other net-
works are less capable of precisely identifying the nodule
0.6 borders leading to a worse Dice concerning GUNET3++.
This is especially true for more complex regions where
the morphology of the nodule’s border is irregular and
0.5
unsymmetrical. This is confirmed, qualitatively, by the
GUNET3++ UNET UNET++ UNET3+
Network example in the second row: the asymmetric nodule
is better identified with respect to the other networks
Fig. 4: Boxplot of dice score distribution of hyper- (UNET++ is unable to identify the irregular boundary).
parameter optimization process for all fine-tuned net- From a quantitative point of view, this is also con-
works. firmed by the distribution analysis of Dice scores for
best-fine-tuned networks with respect to all segmented
nodules. The histogram in Fig. 6 shows the distribution
metrics (Dice, SEN, PPV). Finally, we highlight the best of Dice for the LIDC-IDRI test set. As we can see,
configuration in yellow. The results obtained show a the figure highlights that most of the test instances for
significantly improved value for the average Dice by both UNET++ and UNET3+ are above 0.85, indicating
≈ 3%. The reason behind such improvement is due to the good performance for all those networks with respect to
ability of GUNET3++ of using contextual information at standard UNET that is more concentrated in the range
the same resolution using horizontal skip connections [0.75, 0.85]. It is interesting however to observe that with
jointly with multi-scale operations. This help to reduce respect to both UNET++ and UNET3+, GUNET3++ Dice
the size of the predicted ROI allowing the network to values are more concentrated in the higher narrower in-
better focus on the nodule effectively ignoring redundant terval [0.85, 0.95], indicating better average performance.
information (i.e., other kinds of tissues that are not
relevant). VII. T HREATS TO THE VALIDITY
Figure 4 shows the boxplots of the Dice score of
each HPO process execution for each of the fine-tuned In this study, it is possible to identify three different
networks. As we can see standard UNET is the worst types of threats to validity: construct, internal and ex-
network in terms of Dice average (all configurations ternal. Respectively, for the first, the dataset adopted for
have a Dice lower than ≈ 0.72). It is also interesting the experiment could represent a possible limitation. To
that with respect to other networks, the number of bad overcome this problem, we chose a dataset that has been

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on February 11,2025 at 04:23:51 UTC from IEEE Xplore. Restrictions apply.
 UNET++ UNET3+ GUNET3++
Radiologist Annotation
Dice=0.9022 Dice=0.9102 Dice=0.9343

UNET++ UNET++ GUNET3++

Radiologist Annotation
Dice=0.8971 Dice=0.9022 Dice=0.9129

Fig. 5: Qualitative analysis of predicted ROI.

Network GUNET3++ UNET UNET++ UNET3+ of the approach is affected by the network architectures
we have chosen, and different architectures may lead
to a variation in the results. To mitigate this threat, we
40 evaluated four different network configurations in the
study. Finally, as far as threats to external validity are
Number of nodules

30 concerned, these concern the generalization of results.

The data set considered includes a large number of
patients and related images, but it would be necessary
20
to test the approach on datasets containing images with
different formats, resolutions, and colors.
10

0 VIII. C ONCLUSIONS
60 65 70 75 80 85 90 95 100
Dice coefficient
The problem of lung cancer detection has been ad-
Fig. 6: Best fine-tuned networks histogram of Dice dis- dressed in many recent studies. The centrality of the
tribution over test-set. problem stems from the importance of early lung cancer
detection so that patients have a longer life expectancy.
In this work, a novel approach has been proposed whose
referenced in numerous medical and engineering stud- goal is to detect the presence of lung cancer using CT
ies. Furthermore, a pre-processing and cleaning phase scan images, which are used to train an enhanced ver-
was carried out to filter all images of different formats sion of the UNET segmentation network. The proposed
which had low quality. Finally, a data labeling process approach was validated using the LIDC-IDRI dataset.
was carried out which involved four radiologists. In The proposed network, GUNET3++, is the best per-
addition, some patients were not considered because forming network, compared to the other variants of
there were parsing errors in the metadata, or there UNet, in terms of Dice having a high number of con-
were overlapping errors of different acquisitions. For figurations reaching an average Dice of about 0.96.
internal validity threats, the results could be influenced Furthermore, even considering a qualitative analysis,
by variables not considered. The dataset has been split GUNET3++ is more able to precisely identify the bound-
into training and test sets with a percentage of 80/20, aries of the nodule leading to a better Dice than the other
and we can’t know if different splits would lead to the variants. A possible future development could be the
same conclusions or not. Furthermore, the performance expansion of the dataset used to generalize the results.

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on February 11,2025 at 04:23:51 UTC from IEEE Xplore. Restrictions apply.
 R EFERENCES detection and classification using deep learning combined with
multiple strategies. Sensors, 19(17), 2019.
[1] Pasquale Ardimento, Lerina Aversano, Mario Luca Bernardi, and [18] International Agency For Research on Cancer. Estimated
Marta Cimitile. Deep neural networks ensemble for lung nodule age-standardized mortality rates (World) in 2020, worldwide,
detection on chest ct scans. In 2021 International Joint Conference both sexes, all ages. https://gco.iarc.fr/today/data/factsheets/
on Neural Networks (IJCNN), pages 1–8, 2021. cancers/39-All-cancers-fact-sheet.pdf, 2020. [Online; accessed 01-
[2] Lerina Aversano, Mario Luca Bernardi, Marta Cimitile, Martina February-2021].
Iammarino, Paolo Emidio Macchia, Immacolata Cristina Nettore, [19] International Agency For Research on Cancer. Estimated number
and Chiara Verdone. Thyroid disease treatment prediction with of new cases from 2020 to 2040, Both sexes, age [0-85+] Trachea,
machine learning approaches. In Jaroslaw Watróbski, Wojciech bronchus and lung. https://gco.iarc.fr/tomorrow/en/dataviz/
Salabun, Carlos Toro, Cecilia Zanni-Merk, Robert J. Howlett, and bars?mode=population&cancers=15, 2020. [Online; accessed 01-
Lakhmi C. Jain, editors, Knowledge-Based and Intelligent Information February-2021].
& Engineering Systems: Proceedings of the 25th International Con- [20] World Health Organization. Knowledge into Action
ference KES-2021, Virtual Event / Szczecin, Poland, 8-10 September Cancer Control - WHO Guide for Effective Programmes.
2021, volume 192 of Procedia Computer Science, pages 1031–1040. https://www.who.int/cancer/modules/Early%20Detection%
Elsevier, 2021. 20Module%203.pdf, 2020. [Online; accessed 04-February-2021].
[3] Lerina Aversano, Mario Luca Bernardi, Marta Cimitile, and Ric- [21] Narinder Singh Punn and Sonali Agarwal. Inception u-net archi-
cardo Pecori. Fuzzy neural networks to detect parkinson disease. tecture for semantic segmentation to identify nuclei in microscopy
In 29th IEEE International Conference on Fuzzy Systems, FUZZ-IEEE cell images. ACM Trans. Multimedia Comput. Commun. Appl., 16(1),
2020, Glasgow, UK, July 19-24, 2020, pages 1–8. IEEE, 2020. February 2020.
[4] Lerina Aversano, Mario Luca Bernardi, Marta Cimitile, and Ric- [22] Diego Riquelme and Moulay A. Akhloufi. Deep learning for lung
cardo Pecori. Deep neural networks ensemble to detect COVID-19 cancer nodules detection and classification in ct scans. AI, 1(1):28–
from CT scans. Pattern Recognit., 120:108135, 2021. 67, 2020.
[5] Ujjwal Baid, Sanjay Talbar, Swapnil Rane, Sudeep Gupta, [23] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional
Meenakshi H. Thakur, Aliasgar Moiyadi, Nilesh Sable, Mayuresh networks for biomedical image segmentation. In Medical Image
Akolkar, and Abhishek Mahajan. A novel approach for fully Computing and Computer-Assisted Intervention – MICCAI, 2015.
automatic intra-tumor segmentation with 3d u-net architecture [24] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully con-
for gliomas. Frontiers in Computational Neuroscience, 14:10, 2020. volutional networks for semantic segmentation. IEEE Transactions
[6] Maria G. Baldeon Calisto and S. Lai-Yuen. Adaresu-net: Multiob- on Pattern Analysis and Machine Intelligence, 39(4):640–651, 2017.
jective adaptive convolutional neural network for medical image [25] Wei Shen, Mu Zhou, Feng Yang, Dongdong Yu, Di Dong, Caiyun
segmentation. Neurocomputing, 392:325–340, 2020. Yang, Yali Zang, and Jie Tian. Multi-crop convolutional neural
networks for lung nodule malignancy suspiciousness classifica-
[7] Aaron Carass, Snehashis Roy, Adrian Gherman, Jacob C. Rein-
tion. Pattern Recognition, 61:663–673, 2017.
hold, Andrew Jesson, Tal Arbel, Oskar Maier, Heinz Han-
[26] Run Su, Deyun Zhang, Jinhuai Liu, and Chuandong Cheng.
dels, Mohsen Ghafoorian, Bram Platel, Ariel Birenbaum, Hayit
Msu-net: Multi-scale u-net for 2d medical image segmentation.
Greenspan, Dzung L. Pham, Ciprian M. Crainiceanu, Peter A.
Frontiers in Genetics, 12:140, 2021.
Calabresi, Jerry L. Prince, William R. Gray Roncal, Russell T.
[27] Wenqing Sun, Bin Zheng, and Wei Qian. Automatic feature learn-
Shinohara, and Ipek Oguz. Evaluating white matter lesion seg-
ing using multichannel roi based on deep structured algorithms
mentations with refined sørensen-dice analysis. Scientific reports,
for computerized lung cancer diagnosis. Computers in Biology and
10(1):8242–8242, May 2020. 32427874[pmid].
Medicine, 89:530–539, 2017.
[8] Jason L. Causey, Yuanfang Guan, Wei Dong, Karl Walker, Jake A.
[28] Nima Tajbakhsh and Kenji Suzuki. Comparing two classes of end-
Qualls, Fred Prior, and Xiuzhen Huang. Lung cancer screening
to-end machine-learning models in lung nodule detection and
with low-dose ct scans using a deep learning approach, 2019.
classification. Pattern Recogn., 63(C):476–486, March 2017.
[9] S. Das and S. Majumder. Lung cancer detection using deep
[29] Floris van Beers, Arvid Lindström, Emmanuel Okafor, and
learning network: A comparative analysis. In 2020 Fifth Inter-
Marco A Wiering. Deep neural networks with intersection over
national Conference on Research in Computational Intelligence and
union loss for binary image segmentation. In ICPRAM, 2019.
Communication Networks (ICRCICN), pages 30–35, 2020.
[30] Michael W. Vannier, Ayman El-Baz, Garth M. Beache, Georgy
[10] S. Durga, R. Nag, and E. Daniel. Survey on machine learning Gimel’farb, Kenji Suzuki, Kazunori Okada, Ahmed Elnakib,
and deep learning algorithms used in internet of things (iot) Ahmed Soliman, and Behnoush Abdollahi. Computer-aided
healthcare. In 2019 3rd International Conference on Computing diagnosis systems for lung cancer: Challenges and methodologies.
Methodologies and Communication (ICCMC), pages 1018–1022, 2019. International Journal of Biomedical Imaging, 2013:942353, 2013.
[11] Junfeng Gao, QingZeng Song, Lei Zhao, XingKe Luo, and [31] C. White, T. Flukinger, J. Jeudy, and Joseph J Chen. Use of a
XueChen Dou. Using deep learning for classification of lung computer-aided detection system to detect missed lung cancer at
nodules on computed tomography images. Journal of Healthcare chest radiography. Radiology, 252 1:273–81, 2009.
Engineering, 2017:8314740, 2017. [32] Kun-Hsing Yu, Tsung-Lu Michael Lee, Ming-Hsuan Yen, S C
[12] Ross Gruetzemacher, Ashish Gupta, and D. Paradice. 3d deep Kou, Bruce Rosen, Jung-Hsien Chiang, and Isaac S Kohane.
learning for detecting pulmonary nodules in ct scans. Journal of Reproducible machine learning methods for lung cancer detection
the American Medical Informatics Association, 25:1301–1310, 2018. using computed tomography images: Algorithm development
[13] Muzzamil Javaid, Moazzam Javid, Muhammad Zia Ur Rehman, and validation. J Med Internet Res, 22(8):e16709, Aug 2020.
and Syed Irtiza Ali Shah. A novel approach to cad system for [33] Guobin Zhang, Zhiyong Yang, Li Gong, Shan Jiang, Lu Wang,
the detection of lung nodules in ct images. Computer Methods and Xi Cao, Lin Wei, Hongyun Zhang, and Ziqi Liu. An appraisal of
Programs in Biomedicine, 135:125–139, 2016. nodule diagnosis for lung cancer in ct images. Journal of Medical
[14] Guixia Kang, Kui Liu, Beibei Hou, and Ningbo Zhang. 3d multi- Systems, 43(7):181, 2019.
view convolutional neural networks for lung nodule classifica- [34] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima
tion. PLOS ONE, 12(11):1–21, 11 2017. Tajbakhsh, and Jianming Liang. Unet++: Redesigning skip con-
[15] Chi Li, M. Zeeshan Zia, Quoc-Huy Tran, Xiang Yu, Gregory D. nections to exploit multiscale features in image segmentation.
Hager, and Manmohan Chandraker. Deep supervision with IEEE transactions on medical imaging, 39(6):1856–1867, Dec 2019.
intermediate concepts, 2018. [35] W. Zhu, C. Liu, W. Fan, and X. Xie. Deeplung: Deep 3d dual
[16] Ingo Lütkebohle. LIDC-IDRI - The Cancer Image Archive (TCIA) path nets for automated pulmonary nodule detection and classi-
Public Access. https://wiki.cancerimagingarchive.net/display/ fication. In 2018 IEEE Winter Conference on Applications of Computer
Public/LIDC-IDRI/, 2014. [Online; accessed 30-January-2021]. Vision (WACV), pages 673–681, 2018.
[17] Nasrullah Nasrullah, Jun Sang, Mohammad S. Alam, Muham-
mad Mateen, Bin Cai, and Haibo Hu. Automated lung nodule

Authorized licensed use limited to: Mepco Schlenk Engineering College. Downloaded on February 11,2025 at 04:23:51 UTC from IEEE Xplore. Restrictions apply.