Privacy-Preserving Transfer Learning Framework for Kidney Disease Detection

Abstract

This paper introduces a new privacy-preserving transfer learning framework for the classification of kidney diseases. In the proposed framework, transfer learning is employed for feature extraction, and differential privacy is used to obtain noisy gradients. A variety of CNN architectures, including Xception, ResNet50, InceptionResNetV2, MobileNet, DenseNet201, InceptionV3, and VGG19 are utilized to evaluate the proposed framework. Analysis of a large dataset of 12,400 labeled kidney CT images shows that transfer learning architectures based on the proposed framework achieve excellent accuracy ratios in privacy-preserving classification. These results demonstrate the effectiveness of the proposed framework in enabling transfer learning models to classify kidney diseases while ensuring privacy. The MobileNet architecture stands out for its exceptional performance, with an impressive accuracy of 99.83% in privacy-preserving classification. Considering the findings of this study, it is evident that the proposed framework is appropriate for the early and private diagnosis of kidney diseases and promotes the achievement of promising results in this field.

1. Introduction

Chronic kidney disease (CKD) is a widespread health issue that impacts over 10% of the global population and directly affects an incredible amount of 800 million people. The prevalence of CKD is notably higher in older people, women, racial minorities, and those with chronic conditions like as diabetes mellitus and hypertension. CKD is a prominent contributor to global mortality and is among the limited number of non-communicable diseases that have contributed to a rise in deaths during the past two years [1]. In the year 2016, the disease ranked as the 16th most common cause of mortality. Projections indicate that by 2040, it will rank as one of the most lethal diseases, increasing to the fifth position [2]. The high number of affected individuals and the significant negative impact of CKD call for urgent solutions to develop early diagnosis systems [1].

Machine imaging techniques, including magnetic resonance imaging (MRI), computed tomography (CT), positron emission tomography (PET), mammography, ultrasound, and X-ray, are extensively employed in the medical imaging domain [3]. Recent developments in medical imaging methods have been crucial for the early diagnosis and treatment of diseases. Nevertheless, this procedure is vulnerable to human error as the task of visual interpretation is typically carried out by highly skilled professionals and these professionals are not always available. Therefore, automated systems are always needed to help doctors’ diagnosis decisions.

Although traditional machine learning approaches may not be adequate for properly and effectively analyzing the complexity of medical images, the combination of deep learning methods and the advancement of fast processors enables precise and efficient analysis of medical images. This approach enables the identification and interpretation of patterns in medical imaging. Historically, the medical sector has not completely harnessed artificial intelligence owing to hardware constraints. However, recent advancements in deep learning algorithms have seen significant progress in the healthcare domain. These advancements have resulted in a monumental revolution in the domain of medical image processing. Specifically, considering the constraints of the convolutional neural networks (CNNs) approach, deep learning models show significant promise in the healthcare sector. Utilizing this approach enables healthcare practitioners to examine complex patterns in medical images with higher precision, therefore offering more efficient and expedited solutions in the realms of diagnosis [4].

CNN is a highly efficient deep learning model in the field of image processing, primarily employed for image classification and recognition. A CNN is composed of several layers that divide the inputs into smaller regions, employing convolution filters. Customization of this model enables the identification and determination of significant characteristics in medical images. In the context of CT kidney imaging, a CNN model can provide significant assistance in the diagnosis of kidney diseases [5].

Radiological imaging is of utmost importance in the identification and assessment of kidney problems. These images provide a detailed depiction of the internal structure and operational characteristics of the kidneys. Nevertheless, the interpretation of these images is a laborious and specialized procedure. CNN models have the capability to identify anomalies in CT scans, therefore facilitating the early diagnosis of diseases. The implementation of this technology can facilitate the early diagnosis of diseases and enhance the management of treatment. The advancement and use of these technologies can serve as a crucial measure in addressing significant health issues, including CDK [6,7].

On the other hand, studies emphasize the issue that disease data, such as CT scans, are highly sensitive. Rigid privacy measures are essential in this situation to protect patients’ sensitive information from attacks and breaches. Insufficient measures may result in people being subjected to discrimination, stigmatization, or financial loss if their health information is disclosed. Furthermore, implementing privacy measures can support the establishment of trust between patients and healthcare professionals, thereby promoting early diagnosis without the concern of their confidential health information being disclosed. It is essential to use privacy-preserving methods for extracting important information from such data [8,9].

Current kidney disease detection studies face several key challenges, such as the following:

• Lack of high detection accuracy: since kidney disease is highly likely to result in death, high detection accuracy is very important for early diagnosis. The studies in the literature still do not present higher accuracy;
• Privacy concerns and inference attack issue: kidney disease detection studies employ sensitive data, such as CT scans, which may disclose the privacy of the patients.

In order to overcome these challenges, the proposed framework introduces these solutions:

• Creation of optimal classifier: to obtain high detection accuracy, the proposed framework employs an optimal classifier, including appropriate dense layers, which gives better results;
• Implementation of differential privacy for optimizer: the proposed framework utilizes differential privacy for noisy gradient computing. Hence, we applied differential privacy in an Adam optimizer in order to provide privacy.

The main contributions of this paper are listed below:

• We introduce a novel framework that includes an optimal neural network classifier including appropriate dense layers. The proposed framework provides the best results in the literature.
• Motivated by the privacy issue in kidney disease detection, we present the use of a differentially private Adam optimizer in transfer learning to detect kidney diseases by respecting privacy for the first time.

This study introduces a novel framework that aims to facilitate the detection of chronic kidney disorders, including tumors, cysts, stones, and normal tissue, while respecting privacy. This framework facilitates the deployment of differential privacy with CNN-based architectures for the analysis of CT images of the kidney. The objective of the study is to employ CNN-based architectures to identify anomalies in the kidney by acquiring knowledge of particular patterns and characteristics embedded in the images, while ensuring the privacy of the data is preserved.

The rest of this paper is organized as follows. Section 2 includes a summary of prior studies conducted in the field. Section 3 outlines the materials and methods employed in the study. Section 4 introduces the proposed framework that addresses the handled problem. Section 5 presents the experimental studies conducted in this paper. Finally, Section 6 concludes the research by summarizing the essential findings and their implications.

2. Literature Review

The main purpose of this section is to review the studies on detecting diseases using a CNN-based architecture with CT images. Additionally, it provides a table for future research by comparing accuracy ratios and datasets acquired from various catalogs utilized in the current literature. Below is a summary of research in the literature that investigates the use of these architectures with CT scans for different disease detection.

In [10], the authors assessed the efficiency of deep learning techniques in the task of liver segmentation. Comparative analyses were conducted on two U-Net architectures of varying sizes and a basic CNN model. The U-Net model triggered by the ReLU exhibited a highly successful performance with an F1 score of 97.58% when trained on a 256 × 256 input size. In comparison, the U-Net model trained on a 512 × 512 input size showed a 2.61% reduction in the F1 score. The F1 score values of the U-Net models trained with sigmoid and linear activation functions showed a notable reduction in comparison to the 256 × 256 U-Net model.

In [11], Kayhan introduced a 3D U-Net model that incorporates an early fusion technique for the purpose of automatically segmenting organs on CT scans. The proposed model employs a two-stage early fusion technique that incorporates CT scans with distinct color spaces. High accuracy rates were attained in the tests for large-size organs; however, the degree of success was limited for small-size organs. Training and testing of the proposed model yielded an average segmentation accuracy of 88.8% and 79.6% for all organs, respectively. The automated segmentation procedure produced successful and promising outcomes. However, further inclusion of training and test data could enhance the model’s performance and enable advances in segmenting tiny-size organs. In comparison to CNN and FCN architectures, the proposed approach demonstrated superior performance.

In [12], Katar examined the advancement of a deep learning system for the purpose of diagnosing COVID -19 using chest CT images. The present system comprises two distinct models that possess both classification and semantic segmentation functionalities. This study employed the EfficientNetB2 model as the classification model and achieved a 99.75% accuracy in classifying chest CT scans as either positive or negative for COVID-19. The U-Net model was employed for semantic segmentation, and a ResNet50-based model achieved the highest performance, producing a Jaccard index of 93.76% and a Dice score of 96.61%.

Al-Areqi et al. [13] explored the application of CT scans in the diagnosis of COVID-19 and assessed the efficacy of machine learning and transfer learning models. A COVID-CT dataset, which included 349 positive and 397 negative cases, was employed. Extraction of shape, statistical, and texture information from CT images was followed by classification using different classifiers. Further classification was conducted using transfer learning models. The findings indicate that the XGBoost classifier achieved the highest accuracy of 98.65%, whereas VGG19 and the ResNet50 model achieved an accuracy of 98.04% and 99.62%, respectively.

The utility of image processing technology in the diagnosis of lung cancer was examined in [14]. A dataset comprising CT images of lung, including healthy, malignant, and benign cases, was utilized. This study aimed to develop a lung cancer diagnosis system by leveraging pretrained convolutional neural networks and a novel hybrid model. Empirical studies have revealed that the developed model greatly enhances the precision of lung cancer diagnosis. Although the accuracy rate by deep learning architectures only reached 92%, the proposed approach demonstrated a notable result with an accuracy of 98.3%.

The performance of CNNs for predicting the risk of COVID-19 is assessed in [15]. The study utilized CT scans acquired from clinically diagnosed COVID-19 patients. The collection comprises 349 CT pictures that are positive and 397 CT images that are negative, extracted from 216 people. Analysis and evaluation were conducted on CNN-based architectures VGG16, ResNet18, ResNet50, DenseNet121, DenseNet169, EfficientNetB0, and EfficientNetB1. The use of pretrained weights resulted in a performance enhancement of approximately 10–15% for all networks, with EfficientNetB1 and DenseNet169 achieving an accuracy level above 85%.

In their study [16], Narmada et al. proposed a CNN model for accurate diagnosis of kidney diseases. The researchers utilized a total of 12,510 CT images obtained from distinct hospitals. The images were classified into four distinct categories all pertaining to kidney health. A CNN-based classification model was developed. Sigmoid and ReLU activation functions were employed. Analysis of the model’s performance was conducted using visual representations, such as a confusion matrix and performance table. The statistical analysis revealed that 99.3% of the whole samples in the test set were accurately classified.

The study by Mehr [17] introduces deep learning techniques for diagnosing lung cancer using lung CT images. The study utilized lung CT images obtained from the Data Science Bowl and Kaggle databases. The classification task utilized the AlexNet and GoogleNet designs, resulting in accuracy rates of 95.919% and 96.360%, respectively.

Wang et al. [18] introduced a novel approach to facilitate the identification of liver disease. Here, they provide an enhanced Deeplabv3+ model specifically designed for segmenting the liver in CT scans. The main enhancement of Deeplabv3+ is the substitution of its backbone with ResNet50, which aims to prevent the occurrence of vanishing gradients. Moreover, a channel attention module is incorporated into every ResNet50 residual block, enabling the network to selectively concentrate on significant information. The incorporation of transfer learning in the improved network serves to expedite the merging process and mitigate the potential for overfitting when dealing with a restricted dataset. By attaining an average union over intersection of 95.40% and a Dice coefficient of 94.80%, the proposed technique surpasses U-net and U-net++, demonstrating exceptional accuracy in liver segmentation.

The study of Hossain et al. [6] demonstrates the efficiency of neural networks in the early detection of chronic kidney disease. Classification of kidney CT images was performed using three distinct neural network classification approaches. Initially, the images underwent watershed segmentation procedures. Next, many neural networks, including EAnet, ResNet50, and a proposed CNN model, underwent training using a dataset that is accessible to the public on Kaggle. The test findings indicate that EANet, ResNet50, and the proposed CNN model attained accuracy rates of 83.65%, 87.92%, and 98.66%, respectively on the test set of classification models. The proposed CNN model demonstrates superior sensitivity and specificity, resulting in the highest overall accuracy.

A study conducted by Xu et al. [19] examined deep learning techniques for the detection of lesions in CT scans of COVID-19. This approach, designed to alleviate the workload of physicians in the diagnostic procedure, is founded on the Faster R-CNN architectural framework. In order to achieve more precise outcomes, the data underwent initial clustering using Kmeans++. Subsequently, the Faster R-CNN model was evaluated using two distinct backbone networks, namely VGG16 and ResNet50. Experimental evaluations were conducted on both the raw dataset and the enhanced dataset. The Faster R-CNN model with the VGG16 backbone demonstrated superior performance on the enriched dataset. Using all test data, this model attained a recall of 68.12% and a precision of 65.58%. The percentage of misclassification was observed as 31.88%.

Consequently, the aforementioned studies have shown that deep learning models yield highly effective outcomes in several classification tasks on medical images. CNN is the most favored model type in these experiments. CNNs exert a significant impact in the field of image processing and analysis, making them very well suited for addressing medical image classification problems.

Table 1. Comparison of previous studies.

Study	Main Architecture	Data Type	Focus	Consider Privacy	Dataset	Number of Records	Accuracy (%)
[10]	CNN	CT	Liver	No	Combined Healthy Abdominal Organ Segmentation	2579	97.58
[11]	CNN	CT	Multiple Organs	No	MICCAI 2015 Multi-Atlas Abdomen Labeling Challenge	3000	88.8
[12]	CNN	CT	Chest	No	EFSCH-19	4000	99.75
[13]	CNN	CT	Chest	No	CT Scans for COVID-19 Classification	19,685	99.62
[14]	CNN	CT	Lung	No	IQ-OTH/NCCD	1190	92.69
[15]	CNN	CT	Chest	No	COVID-CT Dataset	746	81
[16]	CNN	CT	Kidney	No	PACS Systems in Dhaka, Bangladesh	12,510	99.36
[17]	CNN	CT	Lung	No	Data Science Bowl and Kaggle Lung CT Scan Images	285,058	96.36
[18]	CNN	CT	Liver	No	3Dircadb Dataset	2074	95.4
[6]	CNN	CT	Kidney	No	PACS Systems in Dhaka, Bangladesh	12,510	98.66
[19]	CNN	CT	Lung	No	Original	263	68.12
This Study	CNN	CT	Kidney	Yes	PACS Systems in Dhaka, Bangladesh	12,400	99.83

presents a comparison of the CNN architecture with several evaluations conducted on CT images. It shows the utilization of CNN models by several researchers in CT classification tasks and provides a summary of classification evaluations conducted on different CT images using deep learning models. Each study focuses on classification problems related to distinct organs or disorders. It can be seen that this study proposes a privacy-aware kidney disease detection for the first time and provides better results.

3. Material and Methods

This section outlines the materials and methods employed in this study.

3.1. CNN

CNN is extensively and efficiently utilized among deep learning models. The concept of CNN was initially proposed by Kunihiko Fukushima in the 1980s. Yann LeCun and his team subsequently integrated CNN with backpropagation theory to accurately identify handwritten numbers and documents. It is highly advantageous for the processing and analysis of visual data, particularly images and videos. The primary advantage of this technology lies in its capacity to acquire hierarchical representations, therefore facilitating a remarkable transformation in the domains of image processing and computer vision through direct learning from pixel data [20]. Medical image processing has seen the extensive application of CNN in recent years. This model demonstrates exceptional performance in tasks including object identification and recognition, image classification, and segmentation [21,22]. The CNN’s capacity to extract significant characteristics and its high performance have established it as a favored option, particularly for the processing of complicated medical images, like CT scans [15].

A CNN is composed of several layers that convert inputs into smaller regions using convolution filters [3]. Multi-layer architecture is employed to identify visual characteristics. These layers are partitioned into two distinct stages: extraction of features and classification. The feature extraction phase consists of an input layer, a convolution layer, and a pooling layer, whereas the classification phase consists of a fully connected layer and an output module [23].

The input data are provided to the network in their raw form, and the resulting resolution directly impacts the network’s performance. Applying filters, the convolution layer derives feature maps from the incoming data. Efficient extraction of visual features requires the use of a specific number of filters. Filters are shifted over the image and matrix multiplications are executed. Summation of the values derived from the multiplications yields the final value. This layer serves the function of identifying patterns in areas of the image that are often invisible to the human visual system. The initial convolution layers extract features at a lower level, whereas the deepest convolution layers extract features at a higher level. The majority of the intensive processing required by the CNN is carried out in this layer [24]. By decreasing the number and size of the network’s parameters, the pooling layer decreases the computational cost and minimizes memory usage. The pooling layer, following the convolution layer, has filters that are convolved with the output of the convolution layer. Within CNNs, three distinct pooling techniques are employed: maximum, minimum, and average pooling [13,25].

The final layer, known as the fully connected layer, aggregates the data by assigning weights and computes probability values for the purpose of classification [3]. Within the fully connected layer, the neurons from the preceding layer are interconnected in their entirety during the flattening phase. In order to make the data from the hidden layers appropriate for the fully connected layer, a flattening procedure is employed. As a consequence of this procedure, the data are transformed into a vector with one dimension. Iterative coefficient operations are executed on this vector in the fully connected layer. The values derived from these operations are activated by running them through the chosen activation function, resulting in the production of the output value. Within the network, the output layer is responsible for comparing the given value with the label. The error is then computed, and the weights are adjusted using the backpropagation algorithm [24].

3.2. Xception

Xception, introduced by Francois Chollet in 2017, consists of 36 layers [26]. This architecture is an advancement of the conventional CNN model and is considered to be pioneering in the field of deep learning for the classification of images. The fundamental component of Xception is the depthwise separable convolution, which effectively separates the conventional convolution process into two distinct stages: depthwise and pointwise convolutions [27]. Depthwise separable convolutions are a computational technique that integrates depthwise convolution, which involves convolving each filter channel with each input channel, and pointwise convolution, which involves convolving the produced output channels with a 1 × 1 filter [13]. With a parameter count of 22.9 million, the model has been trained to accurately classify 1000 different categories on the ImageNet dataset. By default, it accepts RGB images as input and anticipates an input with the number of dimensions (299,299,3). Furthermore, it is capable of processing images with the dimensions (224,224,3). Figure 1 depicts the architecture of Xception, as described in [28].

3.3. ResNet50

Developed in 2015 by He Kaiming, Sun Jian, and Microsoft Research Asia, the ResNet architecture is a deep neural network that is based on CNNs. In order to tackle the issue of vanishing gradients in deep neural networks, the ResNet architecture was developed. An issue arises when gradients reach inadequate magnitude during backpropagation, therefore impeding the network’s ability to train efficiently. The architecture of ResNet50 has 50 layers, which are composed of convolutional layers, pooling layers, and fully connected layers. The model has undergone pretrained training using the ImageNet dataset, an extensive collection of 1.2 million images divided into 1000 distinct categories. In order to minimize a loss function, such as the cross-entropy between the anticipated output and the actual labels, the ResNet50 architecture can be trained using the backpropagation technique and stochastic gradient descent. This extensive training dataset has established this model as the predominant CNN model in computer vision applications. The architecture of ResNet50 is shown in Figure 2 [29].

3.4. InceptionResNetV2

InceptionResNetV2 is derived from the InceptionV3 and ResNet architectures. The computational cost of this architecture is lower than that of Inception and Resnet. This architecture has integrated Inception with connections, resulting in a total stack of 164 layers. Incorporating residual connections, convolution filters not only decrease the duration of training but also mitigate the occurrence of deep network distortions [30,31]. This architecture is illustrated in Figure 3.

3.5. VGG19

VGG19 was introduced in 2015 by the Visual Geometry Group at Oxford University [32]. Its architecture comprises 19 deep layers, including 16 convolution layers, 5 max pooling layers, and 3 fully connected layers. The architecture has undergone training using the ImageNet dataset to accurately classify 1000 distinct categories, with a total of 143.7 million parameters. The model is designed to process RGB images on input with dimensions of 224 × 224 × 3. Both the convolution and pooling filters have dimensions of 3 × 3 and 2 × 2, respectively. The initial two fully connected layers consist of 4096 neurons, while the last layer comprises 1000 neurons and utilizes a SoftMax activation function. Every layer in this architecture employs the ReLU function for activation. The architectural design of VGG19 is seen in Figure 4.

3.6. MobileNet

MobileNet was developed specifically to optimize efficiency and is especially suitable for operation on embedded or mobile devices. It was officially launched in April 2017 [33]. To minimize the number of features in this model, the base layer is constructed using a depth-separable convolutional architecture, and this architecture effectively minimizes the parameter count. The key distinction between the depth-separable convolutions in the MobileNet architecture and conventional convolutions lies in the utilization of distinct layers for 3 × 3 depth convolution and 1 × 1 point convolution rather than a single 3 × 3 convolution layer. In depth-separable convolution, the convolution operation is performed independently for each input channel. Pointwise convolution is the application of convolution operators in a 1 × 1 dimension to produce convergence across channels. Figure 5 depicts the architectural design of MobileNet.

3.7. InceptionV3

InceptionV3 was initially introduced by Szegedy et al. The architecture has a substantial series of convolution and maximum pooling layers. The topology of the network consists of 48 layers. Factorized convolutions enhance the network’s efficiency by minimizing the parameter count. The neural network employs asymmetric convolution processes. In addition to pooling activities, grid size reduction operations are also executed. These actions facilitate the resolution of bottlenecks [34]. Figure 6 indicates the layers of the InceptionV3 architecture.

3.8. DenseNet201

DenseNet201 was introduced by Huang et al. in 2017 as an extension of the ResNet architecture [35]. It consists of 201 layers, starting with a 7 × 7 convolution filter and then performing a 3 × 3 max pooling operation. The layers are succeeded by 196 convolution layers, embedded with 3 mean pooling layers, and result in an output layer with a fully connected layer consisting of 1000 neurons, namely global mean pooling. This architecture has a parameter count of 20.2 million and is designed to receive RGB images with dimensions of 224 × 224 × 3 as input. The filters are organized into four compact blocks. The layers inside the dense blocks are tightly interconnected, with each layer receiving input from all preceding layers and transmitting its output to the subsequent layers. The architecture of DenseNet201 is seen in Figure 7.

3.9. Differential Privacy

Differential privacy (DP), a widely used privacy-preserving method, was initially proposed by Dwork in 2006 [36]. It aims to present a solution for statistical attacks targeting databases. Differential privacy is a resilient method that protects private information by introducing random noises to query responses. It ensures that the distribution of noisy results closely fits to the distribution of the actual results [37]. The fundamental premise of differential privacy is that statistical information should not be significantly impacted by the existence or absence of any record in a database. Thus, it protects private information against attackers with a background knowledge [9].

$(\epsilon , \delta )$-Differential Privacy: the mechanism M is considered to satisfy $(\epsilon , \delta )$-differential privacy if it meets the constraint stated in Equation (1), for every output $S$, among neighboring datasets $D_1$ and $D_2$. The parameters $\epsilon$ and $\delta$ established the degree of privacy. $\epsilon$ is the upper level of impact that an individual’s record can have on public information, whereas $\delta$ denotes the probability that this upper level does not hold true:

$$\mathrm{P}\mathrm{r}[M\left(D_1\right)\in S\le e^{\epsilon }\ast \mathrm{Pr}\left[M\left(D_2\right)\in S\right]+\delta$$

(1)

For a dataset, $D$, the mechanism $M$ that answers query $f$ with Gaussian noise is constructed as in Equation (2):

$$M\left(D\right)=f\left(D\right)+normal(0,{\sigma }^{2}I)$$

(2)

3.10. Evaluation Methods

Evaluating the performance of the proposed framework with transfer learning architectures is an essential process to measure how well the models have acquired the ability to classify test data. The architectures generate predicted labels, which are then compared with the true labels to determine the performance outcomes. This section examines various evaluation methodologies.

a.

Confusion matrix

As a typical practice, the performance evaluation of a classification algorithm involves comparing the actual and predicted values of the classes with the confusion matrix. An analysis of the predicted labels against the actual labels yields four outcomes: true positive (TP), true negative (TN), false positive (FP), and false negative (FN). TP indicates the number of positive class instances that are accurately identified as positive. TN represents the number of instances belonging to negative class instances that are accurately predicted as negative. FP is the count of negative class instances that the algorithm classifies incorrectly as positive. FN represents the number of positive class instances that the algorithm predicts as negative. A confusion matrix for two classes is presented in Figure 8.

b.

Classification metrics

Classification metrics are also used to measure the performance of the models. Metrics such as accuracy, precision, recall, and F1 score are calculated as follows [39].

Accuracy is the ratio of correct predictions to total predictions and is given in Equation (3):

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(3)

Precision is the ratio of true positive class predictions to the sum of true positive and false positive, shown in Equation (4):

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{c}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(4)

Recall is defined as the ratio of correct positive class predictions to the sum of correct positive class predictions and false negative class predictions and is given by Equation (5):

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(5)

The F1 score is defined as the weighted average of precision and sensitivity and is given by Equation (6):

$$\mathrm{F}1 \mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\displaystyle \frac{2\times \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}}$$

(6)

The area under the process characteristic curve, or AUC, is another evaluation metric used in classification problems. The ROC curve plots the TP ratio (TPR) against the FP ratio (FPR). The ROC curve has the TP ratio on the Y-axis and the FP ratio on the X-axis. These ratios are presented in Equation (7) and Equation (8), respectively:

$$\mathrm{T}\mathrm{P}\mathrm{R}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(7)

$$\mathrm{F}\mathrm{P}\mathrm{R}={\displaystyle \frac{\mathrm{F}\mathrm{P}}{\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}}}$$

(8)

3.11. Dataset

The dataset obtained via the Picture Archiving and Communication System from several hospitals in Dhaka, Bangladesh, was employed in this study. Patients with kidney tumors, cysts, stones, and normal tissue were included. Following the established procedure, whole abdomens and urinary tracts were scanned, and both coronal and axial slices from both contrast and non-contrast scans were used. Next, the Dicom images were carefully chosen individually based on the diagnosis. From these images, a Dicom image stack was created by the data providers, which included the specific area of interest for each radiologic finding. Then the information and metadata of each patient were retrieved from the Dicom images and converted to a lossless jpg image format. Following the conversion process, a radiologist and a medical technician reviewed every image finding to verify the truthfulness of the data [40].

The dataset obtained consists of 12,510 distinct CT images. We selected a subset of 12,400 images, allocating 3709 to cysts, 5031 to normal tissue, 1377 to stones, and 2283 to tumors [40]. Figure 9 shows CT kidney images of normal, cyst, stone, and tumor. Normal kidneys are characterized by a smooth cortex and medulla. Kidney cysts appear on CT images as round or oval-shaped lesions, usually dark in color and with sharp edges. Kidney stones appear as dense masses, usually bright white on CT images. The size and location of the stone determines the severity of symptoms. Kidney tumors may appear different in density from normal kidney tissue. The size, shape, and location of the tumor are also important.

4. The Proposed Framework

This paper introduces a novel and comprehensive framework that offers the dual capability of kidney disease detection and privacy protection. The identification of kidney disease is achieved by the use of transfer learning methods, namely Xception, ResNet50, InceptionResnetV2, MobileNet, VGG19, InceptionV3, and DenseNet201, with fine-tuning. These methods are robust algorithms extensively employed for image classification tasks. Furthermore, the differential privacy paradigm is used to ensure data privacy. Differential privacy can be implemented at several stages of any deep learning system [41]. The proposed framework applies differential privacy to the gradients. Therefore, a differentially private Adam (DP-Adam) optimizer is used in the proposed framework, which is indicated in Figure 10.

Gradient perturbation via differential privacy ensures privacy by releasing a noisy gradient at each iteration while yet allowing for post-processing of data [8]. Developing an optimization algorithm that satisfies differential privacy is based on a differentially private stochastic gradient descent in which differential privacy is applied in the gradient computation phase. In [42], it is emphasized that traditional gradient computation is performed according to the Equation (9), in which $g$, $m^{\prime }$, $\mathcal{L}$, $x_i$, and ${\theta }_t$ indicate gradients, number of elements in mini-batch, loss function, element in mini-batch, and position, respectively:

$$g={\displaystyle \frac{1}{m^{\prime }}}{\sum }_{i=1}^{m^{\prime }}{\nabla }_{\theta }\mathcal{L}(x_i,{\theta }_t)$$

(9)

In addition, the update rules in Adam optimizer are given in detail in [42]. After the calculation of gradients, bounded gradients $\overline{g}$ are computed to ensure the gradients are bounded by C:

$$\overline{g}={\displaystyle \frac{g}{max(1,\frac{{‖g‖}_2}{C})}}$$

(10)

Now, it is time to calculate noisy gradients by adding noise picked from a distribution mechanism:

$$\tilde{g}=\overline{g}+z$$

(11)

In Equation (11), z indicates a vector of random variables. Finally, the noisy gradient $\tilde{g}$ satisfies differential privacy.

In this framework, input image is preprocessed and then transfer learning algorithms are applied for feature extraction. Then the features are flattened, and dense layers which consist of neural networks are applied for the classification task. Dense layers consist of three layers containing 64, 32, and 4 neurons with ReLU and Softmax activators. The DP-Adam optimizer algorithm optimizes the parameters of the neural networks by respecting privacy. Then training and validation loss are compared and then this process iterates for N epochs. Finally, the outputs are obtained.

Some details about the proposed framework shown in Figure 10 are given below:

• Image input: the input image is read for processing;
• Image preprocessing: the read image is firstly resized to 100 × 100, normalized, and then split for 70% training and 30% testing;
• Feature extraction: the features of the images are extracted by using a transfer learning algorithm;
• Flattening: the extracted features are flattened to a one-dimensional array;
• Creation of dense layers: three dense layers are used to create a neural network architecture with the number of neurons—64, 32, and 4. ReLU activator is used in the first two layers, and SoftMax activator is employed for the last layer;
• Compiling model: the model is compiled in order to fit the model;
• Differentially private optimizer: the optimizer is made differentially private in order to provide privacy;
• Loss comparison: training and validation loss are compared;
• Iteration: in order to optimize the model weights, the process is iterated;
• Output: the classification results are obtained.

The network structure of the proposed framework comprises three dense layers after the extraction of the features with fine-tuned transfer learning architectures. Dense1, Dense2, and Dense3 layers include 64, 32, and 4 neurons, respectively. The differentially private classification task is performed by adding Gaussian noise to the raw gradients of the Adam optimizer. With the help of this process, noisy gradients are obtained for training of the model.

5. Experimental Studies

Accurate identification and classification of CT kidney images as cysts, normal, stones, and tumors is particularly crucial in the healthcare sector. Effective implementation of this procedure is crucial in guaranteeing that patients obtain an early diagnosis and accurate and fast medical care. Thus, the proposed framework facilitates the use of transfer learning methods which rely on CNN to accurately classify medical images. The employed CT images in experimental studies are classified using both the traditional approach and the privacy-preserved approach with the proposed framework.

The experiments were carried out on a computer equipped with 16 GB of RAM, a Xenon 3.4 GHz processor, and running the Windows 11 operating system. Python 3.11, TensorFlow 2.16, as well as other libraries, including numpy, pandas, opencv, and matplotlib were used.

5.1. Preprocessing of Dataset

The dataset comprised 12,400 CT kidney images. Among these images, 3709 depicted cysts, 5031 depicted normal tissue, 1377 depicted stones, and 2283 depicted tumors. During the data preprocessing stage, the file paths of the images associated with each class were identified and lists of these files were generated. Next, every image was appended to a list and assigned the appropriate class label. To accomplish this, distinct labels were assigned to cyst, normal, stone, and tumor instances, respectively. Within the image processing phase, every image was read and scaled to dimensions of 100 × 100 pixels. Subsequently, the pixel values were standardized to the range of values between 0 and 1. This procedure organizes the digital representation of the images for subsequent processing. It was configured to assign 70% of the dataset for training and the remaining 30% for testing.

5.2. Experimental Results

This section evaluates and compares the performance of seven distinct CNN architectures, namely Xception, ResNet50, InceptionResNetV2, MobileNet, DenseNet201, InceptionV3, and VGG19, in accurately classifying CT kidney images into cyst, normal, stone, and tumor classes.

The architectures underwent training using same training dataset and hyperparameters, and their performance was evaluated using comparable test datasets. Strict consistency was maintained throughout the training procedure to ensure fair comparison of the model performance in both training and testing stages. DP-Adam was utilized as the optimizer for privacy-aware classification, whereas Adam was used for traditional classification tasks. The parameters assigned to the learning rate, batch size, number of epochs, number of microbatch, l2 norm clip, and number of epochs are 0.0001, 40, 100, 40, 1, and 100, respectively. The loss function utilized was categorical cross-entropy, whereas the activation functions applied were ReLU and Softmax. Application of the trained model enabled the generation of predictions on the test dataset, which were then compared with the true labels. Consequently, the classification performance of each model in distinguishing kidney cysts, normal tissues, stones, and tumors was assessed.

The performance results of Xception, ResNet50, InceptionResnetV2, MobileNet, VGG19, InceptionV3, and DenseNet201 architectures were evaluated and described, both with and without privacy modifications. Presented in

Table 2. Traditional classification results (without privacy).

Architecture	Accuracy	Recall	F1	Precision
Xception	0.9965	0.9965	0.9965	0.9965
ResNet50	0.9978	0.9978	0.9978	0.9978
InceptionResNetV2	0.9954	0.9954	0.9954	0.9954
VGG19	1	1	1	1
MobileNet	0.9997	0.9997	0.9997	0.9997
DenseNet201	0.9997	0.9997	0.9997	0.9997
InceptionV3	0.9905	0.9905	0.9905	0.9905

are the traditional classification results, whereas

Table 3. Privacy-aware classification results with the proposed framework (nm = 0.1, $\epsilon$ = 772,577, and $\delta$ = 1 × 10⁻⁵).

Architecture	Accuracy	Recall	F1	Precision
Xception	0.9736	0.9736	0.9734	0.9735
ResNet50	0.7591	0.7591	0.7148	0.7347
InceptionResNetV2	0.9537	0.9537	0.9534	0.9533
VGG19	0.9400	0.9400	0.9381	0.9388
MobileNet	0.9983	0.9983	0.9983	0.9983
DenseNet201	0.9875	0.9875	0.9875	0.9875
InceptionV3	0.9392	0.9392	0.9383	0.9391

,

Table 4. Privacy-aware classification results with the proposed framework (nm = 0.2, $\epsilon$ = 4625, and $\delta$ = 1 × 10⁻⁵).

Architecture	Accuracy	Recall	F1	Precision
Xception	0.9494	0.9494	0.9489	0.9489
ResNet50	0.7051	0.7051	0.6431	0.6925
InceptionResNetV2	0.9223	0.9223	0.9209	0.9210
VGG19	0.9104	0.9102	0.9103	0.9102
MobileNet	0.9911	0.9911	0.9911	0.9911
DenseNet201	0.9848	0.9848	0.9848	0.9848
InceptionV3	0.8997	0.8997	0.8968	0.8998

and

Table 5. Privacy-aware classification results with the proposed framework (nm = 0.3, $\epsilon$ = 290, and $\delta$ = 1 × 10⁻⁵).

Architecture	Accuracy	Recall	F1	Precision
Xception	0.9271	0.9271	0.9249	0.9265
ResNet50	0.6672	0.6672	0.5829	0.6399
InceptionResNetV2	0.9033	0.9035	0.9034	0.9033
VGG19	0.8947	0.8947	0.8947	0.8944
MobileNet	0.9868	0.9868	0.9867	0.9868
DenseNet201	0.9781	0.9781	0.9782	0.9781
InceptionV3	0.8672	0.8672	0.8676	0.8676

shows the test results of the proposed framework with different noise_multiplier (nm) and $\delta$ results. Generally, the value of $\delta$ is determined as 1 × 10⁻⁵ in the literature. Therefore, throughout the experiments, we set $\delta$ to 1 × 10⁻⁵. The images presented in Figure 11, Figure 12, Figure 13, Figure 14, Figure 15 and Figure 16 were created according to the best results obtained, as in Table 3. Figure 11, Figure 12 and Figure 13 provide the confusion matrix for each class, the accuracy rates of the training and test data, and the ROC curves for traditional classification. Given that Table 3 yielded superior outcomes in the studies, Figure 14, Figure 15 and Figure 16 were generated based on the values indicated in Table 3. Presented in Figure 14, Figure 15 and Figure 16 are the confusion matrices for each class, the accuracy rates of the training and test data, and the ROC curves for privacy-aware classification via the proposed framework. Among the architectures included in Table 1, VGG19 outperformed other architectures in traditional classification tasks. Furthermore, alternative algorithms also yielded robust outcomes for classification in terms of accuracy, recall, F-1, and precision.

According to the results in Table 3, Table 4 and Table 5, the proposed framework tests demonstrate that MobileNet achieved a remarkable accuracy ratio of 99.83% in the classification CT images. The best results were obtained with settings of nm = 0.1, $\epsilon$ = 772,577, and $\delta$ = 1 × 10⁻⁵. In addition, the other architectures similarly achieved satisfactory high precision, recall, F-1, and accuracy outcomes. The results demonstrate that the proposed framework facilitates the effective performance of transfer learning algorithms while maintaining privacy aspects. It can be clearly seen that when nm increases and $\epsilon$ decreases, the accuracy decreases.

6. Conclusions and Discussion

Technological advancements in medical imaging techniques have played a vital role in the early detection and management of illnesses. However, this process is vulnerable to human error since the task of visual interpretation is usually performed by professionals, who may not always be easily accessible. Hence, the use of automated technologies is always necessary to assist doctors in making diagnostic interpretations. However, in the literature, it is known that disease data are highly sensitive, and privacy measures should be taken to protect the privacy of patients against disclosure attacks. Disclosure of such kinds of data may negatively affect data owners, such as via discrimination, stigmatization, or financial harm. Hence privacy measures are always required. On the other hand, implementing privacy measures enables trust to be established between patients and healthcare providers. Thus, any useful process over these kinds of data, such as early diagnosis, treatment planning, etc., may be performed easily without considering if private data are disclosed. It is essential to use privacy-preserving measures for extracting important information from such data.

The objective of this study was to provide a novel framework that considers the accurate classification of kidney diseases while ensuring data privacy. In the proposed framework, transfer learning was employed for feature extraction, and differential privacy was used to obtain noisy gradients. An evaluation of the proposed framework’s efficiency was conducted by comparing the performance of seven distinct transfer learning architectures: Xception, ResNet50, InceptionResNetV2, MobileNet, DenseNet201, InceptionV3, and VGG19. Every architecture was assessed using accuracy, recall, F-1 score, and precision criteria. The acquired findings demonstrate the efficiency of each architecture in classification tasks.

The experimental findings indicate that with nm = 0.1, $\epsilon$ = 772,577, and $\delta$ = 1 × 10⁻⁵, the MobileNet model outperformed the other benchmark models in privacy-sensitive diagnosis of kidney diseases, with an accuracy of 99.83%. This model has exceptional performance characterized by excellent accuracy and well-balanced metrics. In addition, it can be seen from Table 3, Table 4 and Table 5 that the accuracy decreased when ε decreased. This is the expected situation in the domain of differential privacy. The results indicate that the proposed framework is applicable for the early detection of kidney diseases and presents promising privacy preserving outcomes in this domain.

While this paper has primarily focused on the detection of kidney disease with respect to privacy, it should be noted that the proposed framework can be generalized for other types of medical imaging and transfer learning tasks. With the help of appropriate preprocessing approaches and training processes, other types of medical imaging data, such as X-rays, MRI, and PET, can be easily used in the proposed framework. In addition, other transfer learning techniques can be easily employed and implemented in the feature extraction process.

In conclusion, the proposed framework facilitated transfer learning architectures to achieve robust performance in classifying CT kidney images while considering data privacy. The use of such a high-performance framework can enhance healthcare quality by facilitating precise, fast, and private disease diagnosis. Thus, the findings demonstrate that deep learning models can be properly employed in medical diagnostic procedures. Implementing CNN models in CT kidney images significantly enhances and simplifies the precise identification of cysts, tumors, stones, and healthy tissues. Improved quality of healthcare services is achieved by guaranteeing prompt and precise treatment for patients.