Renal CT Image Classification Based on Densely Connected

Convolutional Networks

Guangjie Qian

Data Science and Big Data Technology, Changzhou University, Changzhou, Jiangsu, 213164, China

Keywords: DenseNet121, CNN, Image Classification, Kidney Disease.

Abstract: In response to the current situation of the increasing incidence of kidney diseases worldwide, the efficiency

of traditional clinical diagnosis may not be enough to cope with future needs. Compared with traditional

methods of clinical diagnosis, the automatic classification of renal computed tomography (CT) images based

on convolutional neural networks (CNN) in this study has the potential to significantly improve the efficiency

and accuracy of clinical diagnosis. In the paper, the Densely Connected Convolutional Networks 121

(DenseNet121) model is selected for training on 12,446 CT images, which include categories such as kidney

cysts, kidney stones, tumors, and normal tissues. The model training was performed using an early stopping

strategy and multi-cycle validation loss assessment. Subsequently, the model was tested on an independent

test set to achieve an impressive accuracy of 0.9351 and a precision of 0.9393. The experiments conducted in

this study have garnered a good response, and their high accuracy could potentially enhance the efficiency of

clinical diagnosis and provide better safety for patients.

1 INTRODUCTION

Nowadays, the incidence of renal diseases has

become higher due to factors such aspersonal life

habits and deterioration of the external environment

(Zhang et al., 2019). Kidneys are one of the very

important organs in the human body, which are

responsible for maintaining various functions of the

body. Such as metabolism, fluid balance, and

endocrine functions. Therefore, kidney disease may

have a serious impact on health. Therefore, the ability

to accurately and quickly determine kidney health is

crucial for the timely detection, prevention, and

treatment of kidney diseases.

Currently, most of the clinical diagnosis of kidney

health is done manually by testing of CT (Zhang et

al., 2019). CT, as a medical imaging technique,

utilizes X-rays and computer technology to produce

detailed cross-sectional images of the internal

structures of the body (Goldman, 2007). These scans

offer clearer and more detailed images than ordinary

X-rays, enabling doctors to accurately view organs,

blood vessels, and more within the body. Kidney

diseases, including kidney cysts and kidney stones,

are likely to grow in the patient base with the effects

of frequent diseases and an aging population.

Therefore, the efficiency of traditional manual CT

diagnosis may need some image recognition

algorithms to improve in the future.

At the present time, deep learning has been

applied to the problem of kidney CT image

classification, and CNNs are widely used in several

tasks of image processing (Alzu’bi et al., 2022;

Mehedi et al., 2022). Such as VGG16, ResNet,

MobileNetV2, they all play an important role in this

problem.

The dataset is made up of 12,446 distinct entries,

encompassing 3,709 instances of cysts, 5,077 normal

samples, 1,377 cases of stones, and 2,283 occurrences

of tumors (Islam et al., 2022). The study uses deep

learning model DenseNet121, Adam optimization,

and other methods. The main research process is as

follows: firstly, data preprocessing and data

enhancement are carried out on the original data, and

a model is constructed to classify and analyze renal

CT images using the DenseNet121 deep learning

framework. Then the loss rate, accuracy, precision,

recall, and other indexes of the model are tested on an

independent validation set to evaluate the model

(Arulananth et al., 2024; Magboo & Magboo, 2024),

and the classification result graphs of the test are

output at the same time.

This paper is divided into several parts: the first

and current part is the introduction; the second part

outlines the main methodology used in the study,

Qian and G.

Renal CT Image Classiﬁcation Based on Densely Connected Convolutional Networks.

DOI: 10.5220/0013510200004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 115-119

ISBN: 978-989-758-754-2

115

including specific methods such as data processing

and model construction; the third part describes the

results of the study, including the performance and

evaluation of the model; and the last part summarizes

the whole paper.

2 METHOD

The main approaches of this research include data

preprocessing and enhancement, and model

architecture.

2.1 Data Preprocessing and

Augmentation

In this study, data preprocessing and enhancement of

images prior to model construction aim to enhance

the model's generalization capability and ensure its

robustness in real-world applications. (Shorten &

Khoshgoftaar, 2019).

All renal CT image data were normalized using

the 'ImageDataGenerator' method, a technique that

involves scaling the raw pixel values of the images

from the range 0-255 to between 0-1. This step helps

to optimize the stability of the algorithm during

model training and speeds up model convergence.

Data enhancement techniques are also introduced.

These transitions include stochastic rotation (up to 40

degrees), horizontal and vertical translation (up to

20%). These steps increase the dataset's diversity

while enabling the model to learn the image

variations due to operational differences in real

clinical settings. For example, rotation and translation

simulate the different poses of the patient during

scanning, while zooming and shearing allow the

model to recognize kidney structures in images of

different sizes and scales. Horizontal flipping can

further enhance the model's adaptability to changes in

image orientation.

2.2 Modeling

The DenseNet model can effectively mitigate the

problem of vanishing gradients and enhance feature

propagation and reuse by connecting each layer to all

previous layers, drastically reducing parameter

requirements. Its structural design not only improves

the model's training efficiency but also has superior

performance on multiple image recognition tasks.

Figure 1 below illustrates the structure of the

DenseNet model. (Huang et al., 2017).

Figure 1: DenseNet model structure.(Picture credit :Huang

et al., 2017)

Figure 1 shows the initial image processed by one

convolutional layer through three densely connected

blocks. The layers within each densely connected

block receive the outputs of all previous layers as

inputs, which enhances the transfer of features. The

dense blocks are connected through a transition layer

consisting of a convolutional and pooling layer, and

finally the result is output after the features are varied

through a linear layer (Huang et al., 2017).

Therefore, the DenseNet121 model is selected for

CT image classification in the study.

A global average pooling layer is introduced to

reduce the parameters and mitigate overfitting when

constructing the model. To enhance the stability of

model training, a batch normalization layer is added,

and a fully connected layer with the ReLU activation

function is introduced to improve the model's

nonlinear processing capability. Overfitting of the

model is prevented by culling the layer with the scale

set to 0.5. Finally, the output layer contains four

neurons, each corresponding to a kidney CT image

category, and the probability of each category is

produced using the softmax activation function.

During the training process, the Adam optimizer

is used in the study, which has the property of

adaptive learning rate to converge quickly in the early

stage of model training and remain stable in the later

stage (Hospodarskyy et al., 2024; Reyad et al., 2023).

Meanwhile, the classified cross entropy is applied as

a loss function to optimize the probability distribution

of the model output.

To address the category imbalance in the dataset,

weights were calculated and applied for each category

during model training. The weights of different

categories are set before training to prevent the

DAML 2024 - International Conference on Data Analysis and Machine Learning

116

possibility of biasing back to the category with more

samples during training, which improves the

prediction accuracy of the model for the category

with fewer samples. In addition, to prevent overfitting

in model iterations, optimize model performance, and

save resources. Training was performed using the

early discontinue method. When training was

discontinued when the validation loss did not improve

for three consecutive cycles, the model with the

highest validation accuracy was saved using

ModelCheckpoint.

3 RESULT

The data of this study contains four types of kidney

CT images. The study splits the raw data into a

training set and a test set with a 7:3 ratio. Figure 2

illustrates the training process.

(a) (b)

Figure 2: (a) Loss image, (b) Accuracy image, (c) Recall image, (d) Precision image.(Picture credit : Original)

Figure 2 shows the trends of the model's loss,

accuracy, precision, and recall over six training

rounds. Overall, the loss decreases significantly

during model training, the other three values improve,

and the validation loss gradually decreases.

The DenseNet121 model training process

evaluates the performance of the model at each

iteration with four key metrics: loss value, accuracy,

precision, and recall. These metrics are used to fully

reflect the model capability and save the optimal

model at each iteration. The test results are presented

in Table 1.

Renal CT Image Classiﬁcation Based on Densely Connected Convolutional Networks

117

Table 1: Model Evaluation Training Set.

Metric

Value

Evaluation Loss

0.2218

Model Accuracy

0.9658

Prediction Precision

0.9688

Recall Rate

0.9631

Table 1 exhibits the model training results at the

optimal iteration period. It can be seen that the model

training loss is 0.2218. The low loss shows that the

model fits properly on the training data. The model's

three metrics on the training set are about 0.96%,

reflecting that the model accurately performs the

classification task.

The study conducted a comprehensive evaluation

of the model using an independent validation set with

the same metrics as Table 1 to fully reflect the model's

ability to perform on an unknown dataset. The test

results are shown in Table 2.

Table 2: Model Evaluation Test Set.

Metric

Value

Evaluation Loss

0.3163

Model Accuracy

0.9351

Prediction Precision

0.9393

Recall Rate

0.9325

In evaluating the performance of the renal CT

image classification model, the validation set contains

about 5,000 images covering four categories. The

following results were obtained from the test study:

loss value 0.3163, accuracy 0.9351, precision 0.9393,

and recall 0.9325. These results indicate that the

model demonstrates high classification accuracy and

reliability, effectively recognizing and classifying CT

images related to renal diseases in most cases.

In order to demonstrate the model effect in

practical applications more intuitively, the study also

tested the actual classification prediction on renal CT

images, and the results were visualized and

graphically presented. The test images and labels

were extracted from the validation set, and the

categories were tested on randomly selected images.

The tested images and the predicted and actual values

are finally shown for comparison. Figure 3 shows one

of the results of the above presentation.

Figure 3: Model effect test.(Picture credit : Original)

Figure 3 shows four CT images and predicted

results with four categories: normal kidney, cyst,

stone, and tumor. The predicted results are

summarized as follows: The first picture shows the

normal kidney image and predicted result with the

actual label. The label content (Actual: 0, Predicted:

0) indicates that the predicted and actual values are

identical. The results of others also show that the

predicted values are also in perfect agreement with

the actual values, all of which can indicate that the

model is very objective in practical applications.

4 CONCLUSIONS

In this paper, it is investigated how deep learning can

be combined with renal CT image processing with the

aim of improving diagnostic accuracy and efficiency.

In this study, the DenseNet121 model was finally

trained on 12446 CT images containing different

kidney conditions, which performed well on the

training set, with evaluation metrics such as accuracy

and precision above 0.95. The saved optimal model

achieved an accuracy of 93.51% on the independent

test set. In order to better visualize the model's

effectiveness in real life, samples were extracted from

the study to compare the actual and predicted results,

and the predictions were all correct as can be seen in

the results visualization. All these show that the

DenseNet 121 model has a better ability to recognize

kidney CT images. However, to enhance diagnostic

reliability for patients, this technology requires

further refinement and practical application. After

DAML 2024 - International Conference on Data Analysis and Machine Learning

118

all, there are many complex problems encountered in

medical image recognition, and the accuracy should

constantly strive for 100% while avoiding possible

errors such as overfitting.

ACKNOWLEDGEMENTS

The dataset used in the study was obtained from the

CT kidney dataset shared by Slam et al. This dataset

provided valuable support for the study. The

completion of this study would also like to thank the

authors of all the references in this paper, which were

of great help during the study.

REFERENCES

Alzu’bi, D., Abdullah, M., Hmeidi, I., AlAzab, R.,

Gharaibeh, M., El-Heis, M., ... & Abualigah, L. (2022).

Kidney tumor detection and classification based on

deep learning approaches: a new dataset in CT

scans. Journal of Healthcare Engineering, 2022(1),

3861161.

Arulananth, T. S., Prakash, S. W., Ayyasamy, R. K.,

Kavitha, V. P., Kuppusamy, P. G., & Chinnasamy, P.

(2024). Classification of Paediatric Pneumonia Using

Modified DenseNet-121 Deep-Learning Model. IEEE

Access.

Goldman, L. W. (2007). Principles of CT and CT

technology. Journal of nuclear medicine

technology, 35(3), 115-128.

Huang, G., Liu, Z., Van Der Maaten, L., & Weinberger, K.

Q. (2017). Densely connected convolutional networks.

In Proceedings of the IEEE conference on computer

vision and pattern recognition (pp. 4700-4708).

Hospodarskyy, O., Martsenyuk, V., Kukharska, N.,

Hospodarskyy, A., & Sverstiuk, S. (2024).

Understanding the Adam Optimization Algorithm in

Machine Learning.

Islam, M. N., Hasan, M., Hossain, M. K., Alam, M. G. R.,

Uddin, M. Z., & Soylu, A. (2022). Vision transformer

and explainable transfer learning models for auto

detection of kidney cyst, stone and tumor from CT-

radiography. Scientific Reports, 12(1), 1-14.

Magboo, V. P. C., & Magboo, M. S. A. (2024). Batch Size

Selection in Convolutional Neural Networks for

Glaucoma Classification. Procedia Computer Science,

235, 2749-2755.

Mehedi, M. H. K., Haque, E., Radin, S. Y., Rahman, M. A.

U., Reza, M. T., & Alam, M. G. R. (2022, November).

Kidney tumor segmentation and classification using

deep neural network on ct images. In 2022

International Conference on Digital Image Computing:

Techniques and Applications (DICTA) (pp. 1-7). IEEE.

Reyad, M., Sarhan, A. M., & Arafa, M. (2023). A modified

Adam algorithm for deep neural network optimization.

Neural Computing and Applications, 35(23), 17095-

17112.

Shorten, C., & Khoshgoftaar, T. M. (2019). A survey on

image data augmentation for deep learning. Journal of

big data, 6(1), 1-48.

Zhang, H., Chen, Y., Song, Y., Xiong, Z., Yang, Y., & Wu,

Q. J. (2019). Automatic kidney lesion detection for CT

images using morphological cascade convolutional

neural networks. IEEE Access, 7, 83001-83011.

Renal CT Image Classiﬁcation Based on Densely Connected Convolutional Networks

119