Mobile Networks and Applications

https://doi.org/10.1007/s11036-023-02237-0

RESEARCH

Design of Network Medical Image Information Feature Diagnosis

Method Based on Big Data
Wei Li1 · Hui Liu2

Accepted: 23 August 2023

© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2023

Abstract
In the context of "smart healthcare", due to the substantial increase in medical data and patient diagnostic needs, conven-
tional diagnostic methods are gradually unable to meet the current diagnostic requirements. Therefore, a network medical
image information feature diagnosis method based on big data is designed to improve the effect of disease diagnosis. The
convolutional deep belief network is used to extract the information features of the network medical image in the network
medical image. The t-SNE algorithm is used to select the more valuable network medical image information features in the
extracted features. Using stacking to integrate AdaBoost and Bagging algorithm, the disease diagnosis results are obtained.
The artificial bee colony algorithm is used to optimize the weights of the multi-level ensemble learning algorithm to improve
the accuracy of disease diagnosis. In the multi-level ensemble learning algorithm after weight optimization, the selected
network medical image information features are input and the disease diagnosis results are output. Experiments show that
this method can effectively extract the information features of network medical images and accurately diagnose diseases. At
different spatial resolutions of network medical images, the Kappa values of disease diagnosis of this method are high, and
the lowest Kappa value is about 0.875, which means that this method has high disease diagnosis performance.

Keywords Big data · Network medical image · Information feature · Disease diagnosis · Deep belief network · t-SNE
algorithm

1 Introduction is mainly composed of digital hospital and regional health

informatization. The digital hospital relying on medical big
With the extensive development of the application of big data platform and medical artificial intelligence engine is
data and artificial intelligence, hospital construction is grad- the core of smart medicine. Digital hospital emphasizes the
ually developing in the direction of automation, intelligence interconnection of massive medical information and the
and digitalization. Big data application and clinical intel- automation and intelligence of disease diagnosis[2], among
ligent decision support will be an important part of future which the automatic intelligent disease diagnosis is the most
smart hospital construction. By constructing a medical and important in the current digital hospital construction. At
health service system with comprehensive interconnection present, Computer-aided Diagnosis (CAD) system is a rela-
and sharing of information and highly intelligent decision- tively mature scheme to realize intelligent disease diagnosis.
making, smart hospitals realize the deep interaction between It is mainly used to store and manage clinical data, analyze
patients and public medical resources [1]. and label the patient's test results (such as electrocardiogram,
According to the definition of the Healthcare Information magnetic resonance image, etc.), make preliminary judg-
and Management Systems Society (HIMSS), smart hospital ment of disease status and assist doctors in treatment and
diagnosis [3].
* Hui Liu Traditional CAD usually uses several clinical indica-
[email protected] tors summarized by experience or some manual features
obtained by statistical analysis to simply distinguish dis-
1
Medical Library, Peking University, Beijing 100191, China eases, and often has a low degree of automation. Because
2
Institute of Medical Information, Chinese Academy effective clinical indicators must come from formulas or
of Medical Sciences & Peking Union Medical College, algorithms of authoritative medical guidelines, which are
Beijing 100020, China

13
Vol.:(0123456789)
 Mobile Networks and Applications

selected by doctors, and this requires a lot of clinical prac- The convolutional neural network first undergoes multiple
tice and experience; In addition, manual features need to down-sampling and then recovers by up-sampling, which
collect a large number of clinical data first, and then use inevitably loses a lot of detailed information related to the
mathematical statistical analysis methods to obtain them. original image pixels, affecting the effect of disease diag-
In addition to using clinical data, CAD can also analyze nosis. Lali et al. propose a method (RV-EXCALIBER) that
uploaded medical images [4] to assist doctors in diag- utilizes summary-level data from the large Public Exome
nosis or provide guidance for subsequent surgical plans. Sequencing database (gnomAD) as a control and robustly
Although early CAD can complete some simple disease calibrates the rare variant burden to account for the above
diagnosis based on medical images, the general efficiency bias; Calibrated RVGRS capture the overall effect of rare
is low and the false positive rate is too high. Therefore, the variants through a polygenic inheritance model for disease
early diagnosis of medical image disease mainly depends diagnosis. This method can effectively realize disease diag-
on experienced radiologists and oncologists, combined nosis [8]. However, this method has the problems of limited
with the analysis results of CAD, and double/multiple accuracy, low model robustness, and high risk of overfitting.
image review is used to complete, and the degree of auto- Nezu et al. used Logistic regression analysis to screen
mation is also not high [5]. To this end, more automated the risk factors of cerebral small vessel disease with mild
disease diagnosis methods need to be studied. cognitive impairment, and R software was used to establish
For example, Tiwari et al. proposed a new disease diag- a random forest model for the diagnosis of cerebral small
nosis method based on knowledge injection Context-driven vessel disease with mild cognitive impairment. The experi-
(KI-CD) hierarchical reinforcement learning (HRL). We mental results show that the AUC of this method for the
combine current context and disease symptom knowledge diagnosis of cerebral small vessel disease with mild cogni-
to identify candidate diseases and potential symptoms, tive impairment is 0.804, which has high disease diagnostic
and strengthen the agent for intelligent and context-guided efficiency. This method is of great significance for simple
symptom investigation with informative states and addi- and rapid clinical risk analysis of mild cognitive impairment
tional criticisms called learner criticisms; By combining and early development of scientific prevention and treatment
a hierarchical disease classifier (HDC) with a symptom measures [9].
investigation model to alleviate the problem of symptom Although this method can help the diagnosis model to
state sparsity, the accuracy of disease diagnosis is signifi- train better, it is prone to overfitting on the training sam-
cantly improved. The proposed method performs better in ples, which will affect the diagnosis effect in the case of
disease diagnosis on multiple benchmark datasets, which insufficient disease samples. Sharma et al. proposed a Long
firmly establishes the effectiveness of the proposed disease Short-Term Memory (LSTM) -convolutional Neural Net-
diagnosis method [6]. work (Convolutional Neural Network) combining Attention
When the method is only suitable for the diagnosis of dis- mechanism. CNN) short-term disease diagnosis method,
eases with obvious traits, the scope of application is small. which can reduce the loss of historical information and real-
Geisler et al. used the disease-related regions obtained by ize short-term disease diagnosis. The proposed method was
comparison between groups as input, used convolutional compared with the single LSTM model, CNN model and
neural network to extract disease-related features from dif- LSTM-CNN combination model. The Root Mean square
ferent regions of the image, and then used graph convolu- Error (RMSE), Mean Absolute Error (MAE) and Mean
tional network to combine the extracted features to model Absolute percentage error (MAPE) were used to evaluate
the topology structure between regions, and embedded graph the proposed method. The accuracy of disease diagnosis is
pooling operation in the graph convolutional network. Thus, the highest, which verifies the effectiveness and scalability
the intrinsic connection between brain topology and disease of the method [10]. However, this method has the problem
diagnosis tasks can be adaptively learned. The ADNI data of "underfitting" of model training due to the lack of image
set was used to obtain the accuracy, sensitivity and speci- data, which affects the reliability of disease diagnosis. How-
ficity of the method for the diagnosis of Alzheimer's dis- ever, this method can only help doctors diagnose the disease,
ease and its prodrome. The analysis results show that the and most patients will not go to the hospital without symp-
method has achieved 92.5% accuracy, 88.1% sensitivity and toms, so this method lacks advance and real-time perfor-
96.0% specificity in the diagnosis of Alzheimer's disease [7]. mance in terms of disease diagnosis. Health and medical big
However, this method replaces all the fully connected layers data covers the whole life cycle of people from birth to death
in the convolutional neural network with deconvolution to [11], including not only personal health information, disease
upsample the feature maps containing high-level semantic diagnosis and treatment information, but also the fusion of
information after multiple downsampling to restore them drug use, disease prevention and control, health and medical
to the original size, and fits the segmentation mask on the security, food safety, health-related life behaviors and other
feature maps of the original size. kinds of data related to personal health [12]. The application

13
Mobile Networks and Applications

of big data technology in disease diagnosis will improve the Similar to the restricted Boltzmann machine structure,
comprehensiveness of disease diagnosis. Therefore, a net- the basic network structure of Convolutional Deep Belief
work medical image information feature diagnosis method Network (CDBN) is also composed of a visible layer v and
based on big data is designed to improve the accuracy of a hidden layer o, where the hidden layer can be divided into
disease diagnosis. The convolution depth belief network is a convolutional layer and a pooling layer. First, the connec-
used to extract the information features of network medical tion weight between the visible layer and the hidden layer
images. The stacking algorithm is used to integrate the Ada- of CDBN is considered. Similar to the restricted Boltzmann
Boost and Bagging algorithm, and the diagnosis results are machine (RBM), the connection weight between the state of
obtained by using artificial bee colony algorithm to improve the visible layer and the state of the hidden layer is defined
the accuracy of disease diagnosis. as W, and each connection weight is assumed to have a cor-
responding offset. These weights and offset parameters will
affect the information transmission and feature learning in
2 Network Medical Image Information the network. Specifically, suppose that the size of the input
Feature Diagnosis Method network medical image is NV × NV , then the convolution
layer consists of k filters of size NO × NO , and different fil-
2.1 Feature Extraction of Network Medical Image ters are used to scan the network medical image from left
Information Based on Convolutional Depth to right and top to bottom. Suppose that the receptive field
Belief Network is moved with a Stride of S each time. Then each filter
will transform the original image into an NW × NW matrix,
Traditional network medical image information feature NW = NV − NO + 1 when the step size S is 1. In addition to
extraction methods often design different feature descrip- this, each input unit shares a bias term a and the different
tors for different tasks, and the design method of descrip- filters share a bias term bk.
tors determines the nature of extracted network medical Then, an energy function of CDBN is defined to reflect
image information features, which fundamentally affects the relationship between the state of the network and the
the accuracy of disease diagnosis. Manually setting the fea- weight parameters. In CDBN, the energy function can be
ture extraction method of network medical image informa- defined as the energy correlation between the visible layer
tion not only requires rich professional knowledge, but also and the hidden layer. By defining an appropriate energy
its universality and the efficiency of feature extraction are function, the network can extract discriminative features
often related to the quality of algorithm design. Therefore, from the original medical image data. These features can
the deep learning model characterized by highly automated help the network to understand the structure, texture, shape
feature extraction is gradually becoming the main techni- and other information in the image, so as to realize a more
cal means of network medical image information feature. accurate and effective feature extraction process. The expres-
Compared with traditional feature extraction methods, the sion of network energy function is as follows:
advantages of deep learning models are mainly reflected in N O NW K NO NV
their powerful learning ability, which not only includes the ∑ ∑ ∑ ∑ ∑
E(v, o) = − k
oki,j Wr,s vi+r−1,j+s−1 − bk oki,j − vi,j
deep learning ability of network medical image information i,j=1 r,s=1 k=1 i,j=1 i,j=1
features, but also reflects in their highly autonomous learn- (1)
ing mode. Here, vi,j represents the neuron state at row i and column
As an unsupervised probabilistic generative model, deep j in the visual layer. oki,j is the neuron state in row i and col-
belief network composed of multi-layer restricted Boltzmann
umn j of the k-th filter; Wr,sk represents the connection weight
machine can effectively extract the local information features
matrix at row r and column s between the visible layer and
of network medical images [13]. However, due to the neglect
the hidden layer; vi+r−1.j+s−1 represents the neuron state in
of the local invariance of network medical images and too
row i + r − 1 and column j + s − 1 in the visual layer. K
many learning parameters in the process of information fea-
denotes the number of filters.
ture extraction of network medical images, the efficiency
Equation (1) can be simplified as follows:
of its information feature extraction of high-dimensional
network medical images is reduced. Therefore, based on K
∑ ( ) ∑K NO
∑ NV
∑
the deep belief network, this paper uses the advantages of E(v, o) = − ok ⋅ W̃k ∗ v − bk oki,j − vi,j (2)
convolutional neural network in parameter reduction and k=1 k=1 i,j=1 i,j=1

other aspects, and uses the combination of two deep learn- where • represents the element-wise sum of the dot product
ing models to extract the information features of network between matrices. W ̃ k represents the matrix obtained after
medical images, so as to achieve the purpose of improving one horizontal flip and one up and down flip of matrix W,
the accuracy of disease diagnosis [14].

13
 Mobile Networks and Applications

* represents using the former matrix to transform the lat- Accordingly, the probability of pk𝛼 suppression in the k-th
ter matrix in a convolutional fashion. bk represents filter pooling cell can be defined as follows:
coefficient.
� � 1
Similar to deep belief networks, using the property that P pk𝛼 = 0�v = � �
the states of neurons in the same layer are independent of ∑ bk + W̃ k ∗v (5)
1+ e i,j

each other, the following conditional probabilities can be

obtained assuming that the other layer is fixed: By sampling the conditional probability shown in Eq. (3),
� � � � the degree of decline of the parameters is estimated, and
∑ NV � ̃ k �
P oki,j = 1�v = 𝜑 W ∗ v + b k
the training of the convolutional deep belief network is
i,j=1
� � �∑ i,j�
(3) completed iteratively. The trained convolutional deep belief
P vi,j = 1�o = 𝜑
K ̃ k ∗ ok
W network is used to extract the features of network medical
k=1
i,j
( ) image information. The specific steps are as follows:
where P oki,j = 1|v is the conditional probability of o;
( )
P vi,j = 1|o is the conditional probability of v; 𝜑 is the acti- Step 1: Input the network medical image X and determine
vation function; W ̃ k is the matrix of the k-th filter after one the sampling step size.
horizontal flip and one up and down flip; ok denotes the hid- Step 2:( Calculate
) the initial condition probability
den layer of the k-th filter. Q0 = P O||X 0 of the convolutional layer using Eq. (3),
In addition, because the convolutional layer is downsam- where X 0 is the initial network medical image.
pled by pooling, and the process of pooling is irreversible, Step 3: Sample according to Q0 to obtain O0 , the initial
that is, the convolutional layer cannot be estimated through network medical image information feature of the convo-
the pooling layer, so the pooling operation needs to be modi- lutional layer [15].
fied in order to be able to perform probabilistic inference. Step 4: According to the samples and current parame-
Consider the convolutional deep belief network structure, ters, the conditional probability of convolution layer is
that is, the structure composed of only three layers: the vis- estimated, and the weight and bias of the network are
ible layer v, the convolutional layer o and the pooling layer p. updated. Repeat this step until the conditions for the end
Assume that the pooling layer also contains k units, where k of training are met [16].
is the number of filters in the convolution layer, that is, each Step 5: The trained CDBN is used to extract the infor-
convolution result of the original network medical image is mation features of network medical images, and a more
pooled once, and the convolution matrix NO × NO of each accurate and effective feature extraction process is real-
network medical image is divided into multiple dense and ized.
non-overlapping C × C size matrices, and the aggregation is
carried out on these small matrices. Then, the size of each 2.2 Visualization and Selection of Network Medical
matrix after downsampling in the pooling layer is NP × NP Image Feature Information Based on t‑SNE
( NP = CO , where C is a custom small positive integer),
N Algorithm
where the pooling rule is: each small matrix a corresponds
to a neuron pa in the matrix after pooling, where there is CDBN is a deep learning model for feature extraction and
at most one neuron activated in the small matrix a, and pa representation learning. It can automatically learn the dis-
activates if and only if there is an activated neuron in a. This criminant feature representation from the original network
rule reduces the possible values of matrix a and neurons to medical image data. These features have better expressive
C2 + 1 cases (pa is 0 for all neurons inhibition, a total of one ability in high-dimensional space and can be used to distin-
case; pa set to 1 means that there are neurons activated in the guish different types of images [17].T-SNE algorithm is a
convolutional layer, a total of C2 cases). Since the value of C technology of dimensionality reduction and data visualiza-
is generally small, the conditional probability of the pooling tion. It can map high-dimensional features to low-dimen-
layer p can be calculated through the convolutional layer h. sional space, and visualize it as a scatter plot, so as to better
Let{Ba denote the coordinates of all neurons in a, that is, observe the similarities and differences between samples
}
B𝛼 = (i, j) ∶ oi,j ∈ 𝛼 . The network energy function needs [18]. T-SNE algorithm presents better visualization effect
to be changed as follows: in low-dimensional space by maintaining the similarity rela-
tionship in high-dimensional space.
NO (( ) ) NO Therefore, after using CDBN to extract the features of
∑ ∑
E(v, o) = − oki,j W̃ k ∗ v + bk ok − vi,j (4) network medical images, the purpose of applying t-SNE
i,j
i,j
i,j=1 i,j=1
algorithm is to further visualize and select these features.
Through t-SNE visualization, we can observe the clustering
of different types of images in the dimensionality reduction

13
Mobile Networks and Applications

space, so as to better understand the similarities and dif- an optimizable variable to optimize, and the best simula-
ferences between image features [19].Further improve the tion point of the input network medical image information
diagnostic effect of network medical images and our under- features in the low-dimensional space is obtained. The
standing of the relationship between image characteristics. formula is as follows:
The steps are as follows: m m
� �� �
Step 1: Calculate the network medical image information 𝜕G � � 𝜌 − 𝜌̂ zi� − zj�
= (11)
feature joint probability 𝜌 in m-dimensional space. Similarity 𝜕zi� j�=1 i�=1 1 + ‖zi� − zj� ‖2
{conditional probability
} 𝜌i′|j′ and 𝜌j′|i′ between feature points
x1 , x2 , ⋯ , xm pairwise of network medical image informa- Step 5: In order to obtain the minimum G, the original
tion in m-dimensional space: network medical image information features can be iter-
2 2
ated for many times, and the error of network medical
e−‖xi� −xj� ‖ 𝛿j� image information feature selection results can be reduced
𝜌i��j� = ∑ , i� ≠ l (6)
m
e−‖xj� −xl ‖2 𝛿j�2 by constantly adjusting the number of iterations t and the
i�,l=1
learning rate [20].
Step 6: Define the perplexity 𝜃 . According to Step 1,
𝛿i′ is the variance of the Gaussian distribution centered at
2 2
e−‖xi� −xj� ‖ 𝛿i�
𝜌j��i� = ∑ , j� ≠ l (7)
m 2 2
e−‖xi� −xl ‖ 𝛿i� the network medical image information feature point xi′ ,
i�,l=1
and any specific value 𝛿i′ induces a probability distribution
where xi′, xj′, xl represent the i′ -th, j′ -th, and l-th network 𝜌i′ at all other network medical image information feature
medical image information feature points, 𝛿i′ represents the points. This distribution will increase as 𝛿i′ increases. The
variance of the Gaussian distribution centered at the fea- t-SNE algorithm uses the concept of perplexity 𝜃 to find an
ture point xi′ of the network medical image information, 𝛿j′ optimal 𝛿i′ in a binary search manner. 𝜃 is the main param-
represents the variance of the Gaussian distribution cen- eter controlling the fitting, which will affect the complex-
tered at the feature point xj′ of the network medical image ity of Gaussian distribution in high-dimensional space, so
information. it is necessary to constantly adjust the size of 𝜃 , and output
The high-dimensional spatial joint probability can be the optimal dimension reduction results, that is, the net-
expressed as: work medical image information feature selection results.
The adjustment process is:
𝜌i�|j� + 𝜌j�|i�
𝜌= (8)
2m (1) Use 𝜃 to fit Gaussian distribution in high-dimensional
Step 2: Calculate the network medical image informa- space. According to the fitting results, dimensionality
tion feature joint probability 𝜌̂ in low-dimensional space. reduction processing is carried out to convert the high-
Because the t-SNE algorithm uses the t-distribution in the dimensional data set into a low-dimensional representa-
low-dimensional space with 1 degree of freedom, the joint tion.
{ } (2) Evaluate the data after dimensionality reduction, and
distribution 𝜌̂ between the feature points z1 , z2 , ⋯ , zm of
network medical image information in the low-dimensional calculate the performance index—accuracy. Adjust the
space can be expressed as follows: parameter size according to the evaluation results.
(3) If the dimension reduction result is too simple, it is
1
necessary to increase the value of 𝜃 to increase the
1+‖zi� −zj� ‖2
𝜌̂ = ∑m ,g ≠ l (9) complexity and flexibility of Gaussian distribution, so
1
g,l=1 1+‖zg −zl ‖2 that it can better retain the characteristic information
of the original data. If the dimension reduction result
Step 3: Calculate the KL divergence between 𝜌 and 𝜌̂ , is over fitted, it is necessary to reduce the value of 𝜃 to
which is set as the objective function G: reduce the complexity of Gaussian distribution, prevent
m excessive fitting and improve generalization ability.
� �
� 𝜌
G = KL 𝜌‖̂
𝜌 = 𝜌ln (10) (4) Repeat steps 2–4: repeat the dimensionality reduction,
i�,l=1 𝜌̂ evaluation and adjustment process according to the new
parameter 𝜃 until the desired dimensionality reduction
Step 4: The objective function G is used to derive the result is obtained.
low-dimensional expression corresponding to the input (5) When the maximum number of iterations is reached,
network medical image information features, as shown in the adjustment process is ended.
Eq. (11), and the low-dimensional expression is used as

13
 Mobile Networks and Applications

2.3 Methods of Disease Diagnosis Bagging algorithm trains several base classifiers in paral-
lel and combines them to achieve the purpose of ensemble
Ensemble learning is an idea and method that can improve classification. The specific method is to obtain several sam-
several different or identical weak learners into strong learn- ple sets by sampling back from the network medical image
ers through ensemble methods. Therefore, the multi-level information feature samples, train and synthesize all the
ensemble learning algorithm is used to diagnose the disease classifiers respectively to obtain the integrated results, and
according to the network medical image information fea- the final disease diagnosis output is as follows:
tures selected in SubSect. 2.2. Stacking is used to integrate T
∑
Adaboost algorithm and Bagging algorithm in ensemble Y(x) = argmax ̂
y(x) (15)
learning to build a disease diagnosis model with stronger 𝜏
performance and better adaptability.
Adaboost algorithm is a kind of ensemble learning algo- Stacking algorithm is another ensemble method that can
rithm, which can improve a weak learner into a strong integrate multiple heterogeneous classifiers. It first trains
learner by training several weak learners in series. Let the several first-level learners, and then uses the output of the
network medical image information first-level learners as the training input of the second-level
{ feature sample
} data set learners, and uses the original network medical image infor-
selected in SubSect. 2.2 be X = x1 , x2 , ⋯ , x𝜂 , where xi′
represents the i′ -th network medical image information fea- mation feature sample labels as the labels of the second-
ture sample; 𝜂 represents the number of samples; In the ini- level learners to learn. Using stacking algorithm to integrate
tial state, the weight of each sample is 1𝜂 . In each round of AdaBoost and Bagging algorithm, the training data set for
disease diagnosis is prepared, and it is divided into training
training, the weight of the sample with the wrong disease
set and validation set. The basic learners: AdaBoost classi-
diagnosis will be increased, so that the sample will receive
fier and bagging classifier are constructed. The validation set
higher attention in the subsequent disease diagnosis. The
is used to evaluate the constructed AdaBoost and Bagging
weight of each classifier is given by the following
classifier, and their prediction results on the validation set
equation:
are obtained. Build enhanced features of training set data.
1 − r𝜏 Taking the enhanced features and the real tags of the training
𝜔𝜏 = 0.5ln (12) set as inputs, a Meta learner is constructed. Training meta
r𝜏
learners with enhanced features and real tags. By updating
where 𝜔𝜏 represents the decision weight of the 𝜏 -th learner; the parameters of the meta learner, such as gradient descent,
r𝜏 represents the disease diagnosis error rate of this learner. it can learn the best combination rules and weights. The
In addition, in each round of training, the weights of the trained meta learner is used to predict the test data and get
attributes are updated as follows: the final diagnosis result.
In order to improve the disease diagnosis accuracy of
𝜙𝜏 (x)e−𝜔𝜏 y(x)̂y(x) the multi-level ensemble learning algorithm, the artificial
𝜙𝜏+1 (x) = (13)
𝜆𝜏 bee colony algorithm was used to optimize the weights of
the multi-level ensemble learning algorithm. Artificial bee
where, 𝜙𝜏+1 (x) represents the weight distribution of medi-
colony (ABC) algorithm simulates the actual honey gather-
cal image information feature samples of the network in the
ing behavior of bees, including leaders, followers and scouts
next round, 𝜙𝜏 (x) is the weight distribution of the current
by leading the bee to search for the honey source, initial-
sample, 𝜆𝜏 is the normalization factor that makes 𝜙𝜏+1 (x)
izing the honey source, using roulette wheel gambling to
satisfy the distribution condition, y(x) and ̂ y(x) represent
determine the honey source to be collected, searching for
the actual diagnosis results of the network medical image
new honey sources in the neighborhood of the honey source,
information feature samples and the predicted results of the
comparing the fitness of the new honey source and the origi-
current learner.
nal honey source to determine the honey source that needs
After obtaining the trained models of all base learners,
to be further mined. If the number of honey sources mined
the final output of ensemble learning is as follows:
reaches the limit, lead the bee to mine the new honey source.
T
∑ The main steps in the algorithm are as follows:
Y(x) = sign 𝜔𝜏 ̂
y(x) (14)
𝜏 Step 1: Initialize the total number of honey sources M and
where, T represents the number of learners. the mining limit L. Each bee represents the weight of a
The final output is the weighted combination of all learn- multi-level ensemble learning algorithm.
ers' decisions according to their own weights, and finally the Step 2: The following bee uses roulette wheel method to
disease diagnosis result is output by symbol discrimination. select the mining honey source, and the probability of the

13
Mobile Networks and Applications

honey source being selected is the ratio of its fitness to the The dataset consists of 3D CT scan sequences covering
sum of all honey sources fitness, namely: the entire colon and rectum obtained by scanning a cross
section of the abdominal cavity with a 64-slice multirow
F
𝜓̂i = ∑M (16) ultrasound detector. A total of 366 patients with prelimi-
̂i=1 F̂i nary diagnosis of colorectal adenocarcinoma were col-
lected. The time of the first physical examination of these
Here, 𝜓̂i represents the probability that the ̂i -th nectar patients included a total of 10 years from 2010 to 2020.
source is selected, and F̂i represents the fitness of the ̂i Their CT sequences were taken about a month before sur-
-th nectar source, with the disease diagnosis accuracy as gery, and in order to more clearly observe the original
the fitness. information of the tumor, None of the patients had under-
Step 3: The follower bee searches for a new honey source gone preoperative radiotherapy or chemotherapy before
location near the honey source: filming. Each patient's CT sequence consisted of approxi-
( ) mately 400 to 500 sections of the abdominal cavity with
𝛽̂i (𝜇 + 1) = 𝛽̂i (𝜇) + 𝛾 × 𝜅 × 𝛽̂i (𝜇) − 𝛽̂j (𝜇) (17) different sections. Three consecutive sections containing
the mass lesion were masked by the physician using a CAD
where 𝛽̂i (𝜇 + 1) is the new location of the ̂i -th honey system, indicating that the lesion was located in a specific
source; 𝛽̂i (𝜇) is the original location of the ̂i -th honey location of the rectum/colon. Each mask labeled slice is
source; 𝛾 is a random number in [-1,1]. 𝜅 is the learn- regarded as a CT image of the abdominal cavity, so our
ing factor; ̂i ≠ ̂j , 1 ≤ ̂j ≤ M , ̂j is a randomly specified dataset contains 366 × 3 = 1098 CT images, each of which
individual.
records the information of the abdominal pelvis, abdomi-
Step 4: The lead bee changes into a scout bee, randomly
nal tissue, colon, and rectum. We split the training and
generates new solutions in the solution space, and cal-
test sets with a ratio of about 3:1, where the training set
culates the fitness of the new honey source. If the fitness
contains 823 CT images and the test set contains 275 CT
of the new secret element is better than the fitness of the
images. It is worth reminding that, considering the large
original honey source, the new honey source is mined,
morphological differences of the same mass lesion in dif-
and the exploitation degree of the new secret element
ferent sections, each CT image will be regarded as an inde-
is set to 0, and the current optimal solution is recorded.
pendent sample in our experiment, and the lesions with
Otherwise, the original honey source is mined, and the
different differentiation degrees will not be distinguished.
exploitation degree is increased by 1.
Step 5: Analyze whether the algorithm meets the termina-
(2) Breast disease MRI dataset for testing disease diagno-
tion condition, if it does, the best weight is output, and the
sis.
best result of disease diagnosis of multi-level ensemble
learning algorithm is obtained.
All samples in this dataset are axial suppression
T2-weighted MRIs from middle-aged and older women
who underwent non-invasive breast pathology testing
between 2011 and 2020 and were imaged using a four-
3 Experimental Analysis channel phased-coil placed in the potential patient. Each
T2wMRI contains cells and tissue from the largest cross
3.1 Datasets section of the patient's two breasts, and the lesion contours
are jointly drawn by two expert radiologists with 8 to 10
In this paper, two challenging medical image datasets are
years of experience in breast MRI analysis and pathologi-
used to test the feature extraction effect of the proposed
cal diagnosis of breast disease. The margin of the lesion
method and the disease diagnosis effect. This experiment
excluding mass cystic and necrotic tissue and the area it
mainly uses MATLAB to carry out the simulation experi-
covered were also labeled, but only the lesion with the
ment of medical image information feature diagnosis, input
largest spatial scale was actually labeled on T2wMRI. The
the medical image data set into the platform, and the medi-
sample dataset contains a total of 543 patients, all evenly
cal image data set has a clear description below the added
distributed according to the patient's risk of death from
position of the platform, so as to obtain the medical image
breast disease. About three-quarters of the samples are
information Feature selection results and diagnosis results.
assigned to the training set and the rest to the test set. The
They are described as follows:
ratio of positive and negative samples in the training set
and the test set is kept the same.
(1) A colorectal cancer CT dataset for testing feature
extraction.

13
 Mobile Networks and Applications

3.2 Comparison Experiment and Analysis of feature selection experiments were carried out on the net-
work medical image information features. Figure 2 lists part
In the colorectal cancer CT dataset, a colorectal cancer CT of the network medical image information feature selection
image is randomly selected, and the features of the CT image results, and the network medical image information feature
are extracted using the proposed method, and the feature categories are 3 categories.
extraction results are shown in Fig. 1. According to Fig. 2(a), when the perplexity is 5, in the
According to Fig. 1 (a) and Fig. 1 (b), it can be seen that network medical image information feature selection results,
the proposed method can effectively extract image informa- the distance between features of the same type is far and
tion features in the original network of colorectal cancer CT the aggregation ability is poor, and the distance between
images, and clearly depict the edge feature information of features of different classes is close, the overlap is serious,
colorectal cancer CT images. In the extracted image infor- and there is significant boundary confusion. According to
mation features, the appropriate image information features Fig. 2 (b), when the perplexity is 20, in the network medi-
are selected as the input samples of the subsequent disease cal image information feature selection results, the distance
diagnosis method. between features of the same type is close and the aggrega-
In this paper, when the t-SNE algorithm is used for image tion ability is better, while the distance between features of
information feature selection, the perplexity will affect the different classes is far, the overlap is alleviated, and there is
image information feature selection effect. Typically, the a slight boundary confusion. According to Fig. 2 (c), when
perplexity is between 5 and 50. Experimental method was the perplexity is 35, the distance between features of the
used to determine the value of the parameter, and the per- same class is close, and the aggregation ability is strong.
plexity value was selected as 5, 20, 35. A total of 27 times The distance between features of different classes is far, and

Fig. 1  Feature extraction results

of CT images of colorectal can-
cer (a) Original CT image, (b)
Feature extraction results

(a) Original CT image (b) Feature extraction results

1.0 1.0 1.0 Feature 1

Feature 1 Feature 1
Feature 2 Feature 2 Feature 2
0.8 Feature 3 0.8 Feature 3 0.8 Feature 3
Feature 3

Feature 3

Feature 3

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 1.0 0 1.0 0 1.0

1.0 0.8 1.0 0.8 1.0 0.8
0.8 0.6 0.8 0.6 0.8 0.6
0.6 0.6 0.6
Featu 0.4 0.4 Featu 0.4 0.4 2 Featu 0.4 0.4 2
re 1 0.2 r e2 re 1 0.2 re re 1 0.2 re
0.2 atu 0.2 atu 0.2 atu
0 Fe 0 Fe 0 Fe

(a) perplexity value = 5 (b) perplexity value = 20 (c) perplexity value = 35

Fig. 2  Feature selection Results at Different Perplexity Levels (a) perplexity value = 5, (b) perplexity value = 20, (c) perplexity value = 35

13
Mobile Networks and Applications

the aggregation ability is weak. The no-overlap case as well coefficient is used to measure the disease diagnosis accuracy
as the boundary confusion case. It can be seen that with the of the proposed method when the spatial resolution of differ-
increase of perplexity, the boundary confusion phenomenon ent network medical images is used, and the analysis results
between different types of network medical image informa- are shown in Fig. 4.
tion features is less obvious. When the perplexity is 35, there According to Fig. 4, with the improvement of the spatial
is no boundary confusion, indicating that the feature selec- resolution of network medical images, the Kappa value of
tion result under this parameter is the most ideal. the proposed method in the diagnosis of different disease
The proposed method is used to diagnose diseases in the types shows an increasing trend. When the spatial resolution
network medical images in the breast cancer MRI data set. of network medical image is 256 × 256, the lowest Kappa
In the breast cancer MRI data set, seven images of differ- value is about 0.875; when the spatial resolution of network
ent breast diseases are randomly selected, which are acute medical image is 336 × 256, the lowest Kappa value is about
mastitis, breast tuberculosis, granulomatous mastitis, breast 0.923; when the spatial resolution of network medical image
fibroma, early breast cancer, middle breast cancer, and is 512 × 256, the lowest Kappa value is about 0.95. When the
advanced breast cancer. The proposed method is used to spatial resolution of network medical images is different, the
diagnose the seven images of breast diseases, and the results Kappa values of the proposed method in diagnosing different
of disease diagnosis are shown in Fig. 3. types of diseases are between 0.81 and 1.00, indicating that
According to Fig. 3, the proposed method can effectively the proposed method has extremely high disease diagnosis
complete the disease diagnosis according to the information performance, that is, the disease diagnosis accuracy of the
characteristics of network medical images. According to the proposed method is high.
disease diagnosis results, it can be seen that the proposed Fridman test is a test used to determine whether mul-
method only misdiagnosed acute mastitis as granulomatous tiple samples are homogeneous, and through this test,
mastitis, and the disease diagnosis results of other samples whether the final disease diagnosis results of the proposed
were not significantly different from the actual results, indi- method are different from other control methods under the
cating that the disease diagnosis accuracy of the proposed same data conditions can be determined. The hierarchi-
method is high. cal reinforcement learning disease diagnosis method in
Kappa coefficient is a calculated index for measuring the literature [6], the disease diagnosis method of artificial
generalization prediction performance of the method and the intelligence in literature [7], the disease diagnosis method
accuracy of disease diagnosis based on the multi-classifica- of variant genetic risk score in literature [8], the disease
tion problem. The results of Kappa value calculation are usu- diagnosis method of machine learning in literature [9],
ally 0–1, and can be divided into 5 grades according to the and the disease diagnosis method of CNN-LSTM in lit-
method performance and disease diagnosis results: 0.0–0.2 erature [10] are the comparison methods of this method.
is very low disease diagnosis performance. 0.21 ~ 0.40 is Fridman test was used to test the disease diagnosis results
the general disease diagnosis performance; 0.41 ~ 0.60 is of the six methods and the disease diagnosis results of the
moderate disease diagnostic performance; 0.61 ~ 0.80 is the control model. The higher the rank of the Fridman test,
high disease diagnosis performance; The performance of the higher the accuracy of the disease diagnosis method.
0.81 to 1.00 is extremely high for disease diagnosis. Kappa The rank of the Fridman test of the six methods in disease

Advanced 1.00
breast cancer Disease diagnosis results
0.98
Medium term Actual results
0.96
breast cancer
0.94
Early breast
Kappa value

0.92
Disease type

cancer
0.90
Mammary
gland fibroma 0.88

Granulomatous 0.86
mastitis Resolution ratio 256 ×256
0.84
Resolution ratio 336 ×256
Breast 0.82
Resolution ratio 512×256
tuberculosis 0.80
Acute Breast Granulomatous Mammary Early breast Medium Advanced
Acute mastitis tuberculosis mastitis gland fibroma cancer term breast breast cancer
mastitis cancer
1 2 3 4 5 6 7 Disease type
Sample number

Fig. 4  Kappa coefficient analysis results for different network medical

Fig. 3  Disease Diagnosis Results image spatial resolutions

13
 Mobile Networks and Applications

Table 1  Rank analysis results of Friedman test Author Contribution Wei Li contributed to Writing—Original Draft,
Methodology, and Conceptualization; Hui Liu contributed to Concep-
Method Mean Rank tualization and Writing—Review and Editing.
Proposed method 5.18
Declarations
Reference [6] Method 4.05
Reference [7] Method 4.87 Competing Interests The authors declare no competing interests.
Reference [8] Method 4.03
Reference [9] Method 4.72 Data Availability Data will be made available on reasonable request.
Reference [10] Method 4.86
Conflicts of Interest There are no declared conflicts of interest or com-
peting interests that relate to the article’s content.

diagnosis was analyzed, and the analysis results are shown

in Table 1.
According to Table 1, the average rank of disease diag- References
nosis of the proposed method is significantly higher than
that of the other five methods, indicating that the disease 1. Li J, Meng Y, Ma L et al (2022) A federated learning based
diagnosis accuracy of the proposed method is significantly privacy-preserving smart healthcare system. IEEE Trans Industr
better than that of the other five methods. The experimen- Inf 18(3):2021–2031
2. Collins CJ, Yi F, Dayuha R et al (2021) Direct measurement of
tal results show that the proposed method has high accu- ATP7B peptides is highly effective in the diagnosis of Wilson
racy in disease diagnosis. This is because the proposed disease. Gastroenterology 160(7):2367–2384
method is based on the deep belief network, uses convo- 3. Pei X, Fan H, Tang Y (2021) Temporal pyramid attention-based
lutional neural network, and uses a combination of two spatiotemporal fusion model for parkinson’s disease diagnosis
from gait data. IET Signal Proc 15(2):80–87
deep learning models to extract the information features 4. Kugunavar S, Prabhakar CJ (2021) Content-based medical
of network medical images to improve the accuracy of image retrieval using delaunay triangulation segmentation tech-
disease diagnosis. nique. J Inf Technol Res 14(2):48–66
5. Wen C, Liu S, Liu S et al (2023) ACSN: Attention capsule
sampling network for diagnosing COVID-19 based on chest CT
scans. Comput Biol Med 153:106338
6. Tiwari A, Saha S, Bhattacharyya P (2022) A knowledge infused
4 Conclusion context driven dialogue agent for disease diagnosis using hier-
archical reinforcement learning. Knowl-Based Syst 242:108292
7. Geisler EL, Agarwal S, Hallac RR et al (2021) A role for artifi-
In the era of "smart healthcare", thanks to the gradual cial intelligence in the classification of craniofacial anomalies.
improvement of basic medical facilities and the explosive J Craniofacial Surg 32(3):967–969 (Publish Ahead of Print)
growth of medical data, the diagnostic needs of patients are 8. Lali R, Chong M, Omidi A et al (2021) Calibrated rare variant
also gradually increasing. The traditional manual diagnosis genetic risk scores for complex disease prediction using large
exome sequence repositories. Nat Commun 12:5852
method relying on doctors and the semi-automatic method 9. Nezu N, Usui Y, Saito A et al (2021) Machine learning approach
based on "computer first aid diagnosis" system can not meet for intraocular disease prediction based on aqueous humor
the needs of development. Fully automatic intelligent diag- immune mediator profiles. Ophthalmology 128(8):1197–1208
nosis scheme has become the first choice today. However, 10. Sharma D, Xu W (2021) Longitudinal microbiome and machine
learning: a cnn-lstm based neural network model for disease
automatic disease diagnosis that is completely independent prediction. Genet Epidemiol 45(7):788–789
of doctors, especially intelligent disease diagnosis that relies 11. Tapas S, Péricles M, Galdino JV et al (2021) Grammar-based
on medical image analysis, is still a challenging problem. automatic programming for medical data classification: an
Because it is often difficult to cover every step of the diag- experimental study. Artif Intell Rev 54(6):4097–4135
12. Thamsen B, Yevtushenko P, Gundelwein L et al (2021) Syn-
nosis process by fully automatic methods, and the CAD sys- thetic database of aortic morphometry and hemodynamics:
tems of the past have not adopted the more recent intelligent overcoming medical imaging data availability. IEEE Trans Med
technology. In view of these defects, this paper designs a net- Imaging 40(5):1438–1449
work medical image information feature diagnosis method 13. Liu S, Huang S, Wang S et al (2023) Visual tracking in com-
plex scenes: a location fusion mechanism based on the combi-
based on big data to improve the accuracy of disease diagno- nation of multiple visual cognition flows. Information Fusion
sis. In the future, ensemble pruning algorithms can also be 96:281–296
studied to selectively remove some base classifiers from the 14. Liu X, Hou S, Liu S et al (2023) Attention-based multimodal gli-
ensemble, which can not only improve the ensemble effect oma segmentation with multi-attention layers for small-intensity
dissimilarity. J King Saud Univ - Comput Inf Sci 35:183–195
but also reduce the size of the ensemble. 15. Siar M, Teshnehlab M (2022) A combination of feature extraction
methods and deep learning for brain tumour classification. IET
Acknowledgements This work is supported by no foundations. Image Proc 16(2):416–441

13
Mobile Networks and Applications

16. Beram SM, Pallathadka H, Patra I (2022) A machine learning 20. Zhao GW, Wang Y, Yang C (2022) Simulation of Medical Image
based framework for pre-processing and classification of medical Segmentation Based on Attention and Pyramid Fusion. Computer
images. ECS Trans 107(1):7589–7596 Simulation 39(02):184–190
17. Girum KB, Créhange G, Lalande A (2021) Learning with context
feedback loop for robust medical image segmentation. IEEE Trans Publisher's Note Springer Nature remains neutral with regard to
Med Imaging 40(6):1542–1554 jurisdictional claims in published maps and institutional affiliations.
18. Rüdiger S, Madesta F, Nielsen M et al (2021) Multi-scale fully
convolutional neural networks for histopathology image segmen- Springer Nature or its licensor (e.g. a society or other partner) holds
tation: from nuclear aberrations to the global tissue architecture. exclusive rights to this article under a publishing agreement with the
Med Image Anal 70(12):101996–102010 author(s) or other rightsholder(s); author self-archiving of the accepted
19. Karthik B, Ramkumar G (2022) Comparison of feature extrac- manuscript version of this article is solely governed by the terms of
tion technique for segmentation in human iris recognition under such publishing agreement and applicable law.
uncontrolled environment using cnn algorithm with SVM classi-
fier. ECS Trans 107(1):16785–16795

13