In this study, two datasets are used, namely, Iraq-Oncology Teaching Hospital/National Center for Cancer Diseases (IQ-OTH/NCCD) Kareem (57) and CT Kidney dataset Islam et al. (56).

The IQ-OTH/NCCD dataset consists of healthy and unhealthy subjects suffering from lung cancer in various stages. This dataset can also be accessed from the Kaggle website Kareem (57). In 2019, the data were gathered for more than 3 months. All of the slides were annotated by the radiologists and oncologists of these two centers. The dataset contains 1,190 images featuring CT scan slices from 110 different instances. The dataset is divided into three categories: MNoB. There are 55 normal cases, 40 malignant cases, and 15 benign cases. Digital Imaging and Communications in Medicine (DICOM) format was used to originally gather the images. However, the JPEG format was subsequently included by the IQ-OTH/NCCD dataset itself. Siemens SOMATOM scanner is employed. One millimeter thick slices are used. The dataset is approved by the institutional review boards of the participating hospitals. Each CT scan is composed of 80 to 200 distinct slices. Every slice is a representation of a distinct angle and side of the human chest. For the 40 malignant instances, there are a total of 561 CT images. The majority of the subjects are from Iraq’s middle area. Figure 1 shows sample images of each MNoB class.

The NCTS dataset was gathered from Dhaka hospital, Bangladesh. Proper consent was taken from all the subjects before collecting the data, which was then approved by Dhaka Central International Medical College and Hospital. The dataset consists of 12,446 total abdomen and urogram CT images, where the axial and coronal cuts were taken. Out of this, the number of NCTS images is 5077, 3709, 2283, and 1377, respectively. The dataset can be accessed from the Kaggle website Islam and Mehedi (58). The dataset was originally in DICOM format, which was later converted to JPEG images through the Sante Dicom editor tool. The CT images were annotated by the Philips IntelliSpace Portal application, which was re-verified by a physician and medical technician to avoid any incorrect annotations. Figure 2 shows sample images of each NCTS class.

The proposed modified Xception model, short for the “modified Extreme Inception” model, mainly consists of transferring the weights of the “imagenet” to the target network through transfer learning followed by the fine-tuned network, as shown in figure effig:3. The target network is the modified Xception architecture. In the beginning, the input image is first resized from 512 × 512 to 224 × 224 during pre-processing for reduced complexity. Then, there are mainly three stages, namely, STAGE-I, STAGE-II, and STAGE-III. First, the source “imagenet” dataset is fed to the original XceptionNet, which is passed over 36 convolutional (conv) layers. Out of these, the number of regular conv and depth-wise Separable Conv 2D layers (Sep Conv 2D) is 2 and 34, respectively. The first block consists of two conv layers of 3 × 3 kernels, followed by a Batch Normalization (BN) and Rectified Linear Unit (ReLU) activation function. All blocks, except Blocks 1 and 14, contain a linear stack of residually connected Sep Conv 2D layers. The input (I/P) and output (O/P) dimensions determine whether the residual connections are identity or convolution blocks. When the I/P and O/P dimensions are identical, identity mapping is performed, as shown in Block 5 of Figure 3 . Otherwise, a linear projection is executed using short connections to make the dimensions match each other. Blocks 2, 3, 4, and 13 have a linear projection of conv filter of 1 × 1 kernel. These blocks also undergo max-pooling operations to extract sharp and smooth features and lower the computing cost by decreasing the amount of parameters that must be learned. BN is included after every conv 2D and Sep Conv 2D layers circumvent the local minima issue by translating the activations to the zero mean and unit variance, hence allowing larger gradient steps for faster convergence Ioffe and Szegedy (59). The use of Sep Conv 2D overcomes the limitation of CNN by segregating the regular conv operations into depth-wise/spatial conv and sequential point-wise conv. This results in fewer parameters compared with regular conv and hence reduces the chances of overfitting Chollet (60). Block 5 is repeated five times having identity mappings. The last block 14 is fed to the softmax layer which finally classifies a thousand classes of imagenet dataset such as kite, bulbul, candle, corn, and so on.

Let the kernel “W” of size 3 × 3 be convolved with the I/P image “j(x, y)” and the output “K(x, y)” is represented by Equation (1), which is then fed to a ReLU activation function “ γ ( K ) “, given by Equation (2). Similarly, the next Conv 2D outputs “l(x, y),” given by Equation (3) and then fed to ReLU activation function, given by Equation (4).

Suppose the I/P dimension of the imagenet dataset be of size D p × D p × C , where D p × D p is the I/P image size and C is the number of channels. Let the number of filters/kernels be D of size D q × D q × C .

For a regular conv, the output size will be D r × D r × D . Let the number of multiplications per conv operation be n. Then the size of the filter will be D q × D q × C . As D filters are present and every filter slides horizontally and vertically D r times, the overall number of multiplications is given by Equation (5).

As Sep Conv 2D is combination of both depth-wise convolutions (d) and point-wise convolutions (p), let the number of “d” multiplications be D and “p” multiplications be P. Thus, the overall multiplications for Sep Conv 2D be given by Equation (6).

The dimension of filters for depth wise operations will be of size D q × D q × 1 . Considering C channels for the I/P data, the number of such filters needed are “C.” Therefore, the output will be of dimension D r × D r × C . Considering all “C” channels, the total multiplications (D) is given by Equation (7).

In point-wise operation, “C” channels are subjected to 1 × 1 convolution. Consequently, the filter dimension after this operation will be 1 × 1 × C . For “D” such filters, the output dimension will be D r × D r × D . For point-wise convolution, there will be 1 × C multiplications. Equation (8) provides the total number of point-wise convolution multiplications.

Thus, the overall multiplications for Sep Conv 2D is given by Equation (9).

The relation between the complexity of regular conv and Sep Conv 2D is given by Equation (10).

Let Q(L) be the desired mapping and the original mapping be indicated by Q(L)+L, executed through shortcut connection followed by addition, as depicted in Figure 3 .

Since in Block 2, I/P and output dimensions, namely, L and Q, are not equal, their dimensions are matched through a linear projection S A using a shortcut connection. This is shown in Equation (11).

where Q ( L , { S e } ) depicts the residual mapping and S e is the weight layer. Ignoring BN for simplicity, Q = γ ( S 1 γ ( S 0 ) ) where γ indicates ReLU.

Similarly, the outputs of Block 3 and 4 are represented by Equations (12) and (13), respectively.

Now, the I/P and O/P dimensions from Block 5 to Block 12 are same. Hence, identity mapping is performed by shortcut connections, as shown in Equation (14). In this case, Q = S 2 ( γ ( S 1 γ ( S 0 L ) ) .

The output of Block 13 is given by Equation (15).

Now, the weights of the imagenet dataset trained on the original XceptionNet is transferred to the target network, which consists of customized top layers and fine-tuned blocks, thus assembling a transfer learning and fine-tuned based modified Xception model. As a result of transfer learning, the input CT kidney dataset and lung IQ-OTH/NCCD dataset fed through XceptionNet acts as feature extractors and the Block 14 is fed to our own customized top layers, which consists of Global Average Pooling (GAP) layers, Dropout (DROP) layer having dropout ratio of 0.05, 4, and 3 dense layers for NCTS and MNoB, respectively. Last, a softmax activation function is applied. Then, fine-tuning is performed by unfreezing the top 20 layers leaving BN layers frozen. If BN layers are set as trainable, the first epoch following unfreezing will result in a considerable reduction in accuracy. Thus, an efficient modified Xception approach is proposed that ensures higher performance in effective multi-class classification of NCTS and MNoB CT images. Table 1 explains the proposed architecture in detail.

The NCTS and MNoB CT images are fed as input to the proposed model, which yields multi-class classified output. The input image shape is set as 224 × 224 for both the cases. The batch size, learning rate (lr) and dense layers are set as 12, 0.00001 and 4, respectively. First, all the input image is appended from the NCTS and MNoB image directory. Subsequently, the labels for each of the case are also appended. Since there are four classes in NCTS and three in MNoB, the labels for NCTS ( l N C T S ) range from 0 to 3 and l M N o B range from 0 to 2, respectively. Second, the train test split is maintained in the ratio of 80 to 20%. Third, the Xception model is loaded with “imagenet” weights, and the topmost layers are replaced with GAP, DROP, and four dense layers. Then, fine-tuning is accomplished by defrosting the upper 20 layers while freezing the BN layers. If BN layers are set to trainable, the first epoch after unfreezing will result in a significant accuracy decrease. Later, the modified Xception model is compiled using Adam optimizer, sparse categorical cross-entropy loss function and accuracy metric. This process leads to a multiclassified NCTS and MNoB output. The Algorithm 1 describes the pseudocode of the proposed model.

Google colab pro plus was used for executing the proposed model, whose specifications are listed below. The python, keras and tensorflow version used are 3.8.10, 2.9.0 and 2.9.2, respectively. The GPU “NVIDIA A100-SXM” of 11.6 CUDA version and 83.48 GB RAM are used.

The I/P RGB images of shape 224 × 224 × 3 were divided into 80% training and 20% testing data. An experiment consisting of four networks, namely, Inception ResNet V2, Inception V3, NASNet and the proposed network, was conducted to get the bestNCTS model. The proposed model was trained with a batch size of 12 and an “Adam” optimizer with a learning rate equal to 0.00001 was used. The training was terminated via an early ending callback if the validation loss does not improve after nine epochs. The topmost layers of these networks were replaced with customized layers. Except the proposed model, all other networks were replaced with top layers having 2,048, 1024, and 512 dense layers followed by the ReLU activation function and four dense layers followed by the softmax activation function. Whereas, the proposed modified Xception model was replaced with GAP, DROP of 50% and four dense layers followed by a softmax activation function.

Similarly, an experiment was conducted for MNoB multi-class classification among the proposed model, Inception ResNet V2, Inception V3 and MobileNet V3 Small. The batch size, optimizer, learning rate, early stopping callback criteria and top layers all are kept the same as the NCTS case except three dense layers are used for MNoB classification instead of four.

The performance metrics such as Accuracy (Train, Test, and Validation), Precision, Recall and F1 score were used for evaluating the proposed modified Xception model and other pre-trained networks. Accuracy (Acc) describes closeness between the positively anticipated value and the actual samples, represented by Equation (16). Precision (Pr) is the percentage of correctly anticipated positives. It is given by Equation (17). Recall (Re) determines the fraction of anticipated positives that are accurate, givenby Equation (18). F1-score is the weighted harmonic mean of Re and Pr, represented by Equation (19).

where TruPos, TruNeg, FalPos, and FalNeg represent True Positives, True Negatives, False Positives, and False Negatives, respectively, and Acc represents Accuracy.

Table 2 shows the Train Acc, Validation Acc, and Test Acc for different models, such as Inception ResNet V2, Inception V3, NASNet Large, and the proposed modified Xception model. Inception ResNet V2 achieved least training, validation and testing accuracy of 40.79, 40.81, and 40.81%, respectively. Whereas, the proposed Xception model attained the maximum Train, Validation and Test Acc of 99.79, 99.92, and 100%, respectively.

Table 3 shows the performance metrics of different models for NCTS classification. Inception ResNet V2 obtained 0% Pr, Re, and F1-score for normal, tumor, and stone classes whereas, it obtained 41% precision, 100% recall and 58% F1-score for kidney cyst class. Thus, it means that it is inefficient in distinguishing among normal, tumor, and stone classes. On the otherhand, NASNet Large and Inception V3 outperformed the Inception ResNet V2 model. However, the proposed modified Xception model obtained the best results by achieving 100% Pr, Re, and F1-score for each of the NCTS classes.

Figure 4 depicts the accuracy and loss curves for all the four different models mentioned above. Inception ResNet V2 showed the poorest accuracy and loss curves, showing accuracy of approximately 40%. The Inception V3 and NASNet Large showed training and validation accuracy of less than 100% and training and validation loss above 0%. However, the proposed model achieved the Train and Validation Acc of 100% and loss of 0%. Also, the epochs for the various NCTS models are not constant because an early termination callback was triggered if the validation loss did not decrease after 9 iterations of training.

Figure 5 depicts confusion matrix for each model. Since the Inception ResNet V2 achieved the poorest results, thus, it is unable to detect any other class except kidney cyst. Moreover, it misclassified other classes and considered them to be part of cyst class, whereas NASNet Large showed less misclassification classes than Inception V3. However, the proposed model showed no misclassification results and perfectly classified NCTS classes. Therefore, the proposed model is superior to the existing models for multi-class classification of NCTS classes.

Table 4 depicts the Train, Validation, and Test Acc of different models such as Inception ResNet V2, Inception V3, MobileNet V3 Small, and the proposed model for MNoB classes. Inception ResNet V2 obtained the least Train, Validation and Test Acc of 51.18, 50.91, and 50.91%, respectively. Although Inception V3 and MobileNet V3 Small achieved almost same training and validation accuracy, their testing accuracies are different. MobileNet achieved approximately 2% lower test accuracy than Inception V3, whereas the proposed modified Xception model obtained the highest Train, Validation, and Test Acc of 100, 99.39, and 99.39%, respectively. Table 5 shows the evaluation metrics such as Pr, Re and F1 score of the above mentioned models for each MNoB class. Inception ResNet V2 obtained 51% Pr, 100% Re and 67% F1-score for normal class. It is unable to detect malignant and benign class, thus, proving it to be inefficient. Inception V3 and MobileNet V3 Small performed better than Inception ResNet V2, whereas the proposed model obtained the highest Pr, Re, and F1-score of 100, 100, 98, 94, 100, 100, 97, 100, and 99%, respectively for each MNoB class.

Figure 6 shows the Train Acc, Validation Acc, Train loss and Validation loss curves of the different models and the proposed model for MNoB classes. Inception ResNet V2 showed the poorest accuracy and loss curves, showing accuracy of approximately 50% and sparse categorical cross entropy loss of 0.9543. Inception V3 and MobileNet V3 fits the training data too closely but unable to generalize well on testing data, thus showing the sign of overfitting. However, the proposed Xception model obtained the highest Train Acc of 100% and sparse categorical cross entropy loss of 0.0015. Thus, the training loss curves obtained by NCTS model is better than the MNoB because of the reduced gap between the training and validation curves and, hence, it validates that there are no overfitting cases.

Figure 7 depicts the confusion matrix for each of the MNoB models. Since the Inception ResNet V2 achieved the poorest results, thus, it is unable to detect any other class except normal. Moreover, it misclassified other classes and considered them to be part of normal class. Whereas Inception V3 showed less misclassification classes than MobileNet V3 small. However, the proposed model showed only one misclassification result by considering malignant class to be benign. Otherwise, it perfectly classified MNoB classes. Therefore, the proposed model is superior to existing models for multi-class classification of MNoB classes.

Figure 8 compares the proposed improved Xception model’s computing time to that of the original Xception model in minutes. It is observed that the original Xception network took 14 min to train itself. Whereas, the proposed model took 12 min to train itself. Hence, fine tuning helped the proposed model in faster training and hence, reduced its computational complexity.

Table 6 compares the proposed modified Xception model with the original Xception model in terms of performance metrics for both NCTS and MNoB cases. In NCTS case, the original model obtained 94.75% Average (Avg) precision, 97.5% Avg recall, 95.75% Avg F1 score and 97.07% test accuracy. On the other hand, the proposed model attained 100% of each of the performance metrics value. In MNoB case, the original Xception model obtained 58.33% Avg Pr, 65.33% Avg Re, 61.33% Avg F1 score and 87.27% Test Acc, whereas the proposed model obtained 99.33% Avg Pr, 98% Avg Re, 98.67% Avg F1 score and 99.39% Test Acc. Thus, it can be inferred that the original Xception model performed poorer for MNoB case compared with NCTS case, whereas, the proposed model performed equally good on both NCTS and MNoB cases. Hence, the proposed Xception model is appropriate for renal disease classification.

Figure 9 compares the ROC curve of both NCTS and MNoB models. Each class of NCTS and MNoB models obtained 100% AUC values. Thus, it is evident that both NCTS and MNoB models performed equally good in terms of ROC curve.

The proposed NCTS model is compared with External Attention Transformer (EANET) Islam et al. (56), Compact Convolutional Transformer (CCT) Islam et al. (56), Swin Transformer Islam et al. (56), VGG 16 Islam et al. (56), Inception V3 Islam et al. (56), ResNet 50 Islam et al. (56), DenseNet 201 Qadir and Abd (55), Ensemble DNN Sudharson and Kokil (52), VGG 19 Asif et al. (53) and CNN Narmada et al. (54), as shown in Table 7 . The proposed model surpassed all the state-of-the-art techniques by achieving the highest Acc, Avg Pr, Avg Re and Avg F1-score of 100% each. Hence, the proposed model is felicitous for automatic CKD diagnosis.

Table 8 shows the comparison among the proposed MNoB model and the state-of-the-art models such as Advanced CNN Reddy et al. (37), Advanced CNN + Synthetic Minority Oversampling Technique SMOTE) Reddy et al. (37), Xception Reddy et al. (37) and ResNet 50 Reddy et al. (37), SVM Kareem et al. (38), DenseNet 121 Bhattacharjee et al. (39) and NASNet Large Bhattacharjee et al. (39). ResNet 50 Reddy et al. (37) performed the poorest followed by Xception Reddy et al. (37). However, the proposed modified MNoB Xception model outperformed all the state-of-the-art techniques by achieving the highest Acc, Avg Pr, Avg Re, and Avg F1-score of 99.39, 99.33, 98, and 98.67%, respectively.

Despite encouraging results, the proposed MNoB model could not outperform the NCTS model due to the lack of training data. This issue can be solved by implementing data augmentation techniques to expand the training data. Moreover, since the proposed MNoB and NCTS models are two-dimensional, it cannot extract context from adjacent slices. This problem can be solved by introducing a 3D classification model that can take advantage of inter-slice context and, hence, enhance model’s performance.