Adaptive 3D Trans-ResUnet and Multiscale Trans-RAN for Enhanced Kidney Tumor Segmentation and Classification

At present, the globe is suffering from the several categories of kidney disorders [1]. The kidney disorders are required to be detected at the earliest to elevate the survival ratio of the individuals. In the past decades, the doctors explained the images of "Computed Tomography (CT)" according to their experience of the kidney disorder diagnosis [2]. But, with the technological advancement in the CT imaging, analysis, and explanation prerequisite more effort and time, and creating inconsistent outcomes. Several frameworks under the Neural Network (NN) technique were proposed to automatically detect kidney tumour locations from CT images [3]. In some research works, the NN frameworks were only modified to enhance the correctness. The accurate segmentation of the kidney tumor in the healthcare images like CT scan images is the primary factor for the better treatment [4]. Different stages of kidney renal cancer by using various algorithms were being used and the success rate was an important parameter to determine the effectiveness of the approaches. The segmentation approach thus plays an important role in establishing the connection amongst the kidney tumor and its appropriate surgical solution and supports the doctors in forming multiple corrective plans for the treatment [5]. However, the manual tumor segmentation takes a significantly more time. The demand for highly correct kidney tumor identification components are hence required [6].

In the modern days, the deep learning approaches have been developed as a promising solution for the medical image segmentation and tumor detection [7]. Among these approaches the Convolutional Neural Network (CNN) framework has outperformed the conventional approaches significantly in the CT image segmentation [8]. The Fully Convolutional Network (FCN) framework has tremendous potential to offer the end-to-end segmentation solution [9]. This kind of framework was also capable to generate the pixel level, raw scale labels in the picture and widely used in the various sectors [10]. Other modern experiments have considered FCN as an initial point toward the more complex and deeper segmentation frameworks like "SegNet". The high-resolution CT images may offer the essential anatomical descriptions to monitor the progression of a disorder like Alzheimer’s disorder in the applications of neural computing [11]. But, explaining these pictures needs a significant amount of effort and time to recognize the regions of the tumor manually [12].

Thus, the deep learning approaches were used for better kidney tumor detection and segmentation to continuously detect and observe the regions of kidney tumors [13]. These approaches allow precise and automatic segmentation of the kidney tumor, hence makes it easier to perform the therapy and diagnosis [14]. Moreover, the approaches can enhance the time and labour efficacy. Eventually, these approaches offer highly corrective and consistent detection outcomes [15]. Presently, the approaches of deep learning have been utilized successfully in multiple sectors employing the physiological signals and the medical images [16]. The deep mechanisms have also been utilized in multiple sectors like medical image segmentation, tumor detection, and categorization [17]. Multiple kinds of images have been employed to enhance the robust and accurate deep-learning approaches to help the clinicians in the detection of diseases such as kidney tumors.

The existing work carried out in the domain of Kidney tumor detection is presented in the following subsection.

1.1 Related works

Hsiao et al. [18] have addressed the necessity of the pre-processed approaches in the NN algorithms. The numerical solutions displayed that the suggested approaches highly enhanced the rate of accuracy contrasted with the approaches which were not involved with the pre-processing. Moreover, the score of dice was also increased for the kidney tumor detection. An orderly evaluation of the process of validating and selecting training data can improve the performance of a model, resulting in consistent segmentation results that can be used for healthcare improvements with little computing needs. The effectiveness and the cost efficacy were also attained for the automatic kidney tumor identification approach.

da Cruz et al. [19] have presented an experiment that was suggested to help the clinicians specialized in the CT kidney tumor detection. Initially, the pre-processing stage was carried out for the standard data-sets and the segmentation was performed. Finally, utilizing the image processing the false positive reduction was accomplished. The resultant outcome reported better accuracy. In 2020, Zhao et al. [20] have deployed a multi-scale approach to segment the kidney tumors from the images of CT. Experts presented the post-processing approach to increase the overall approach functionality. The framework illustrated the superior functionality contrasted to the other promising tasks with few of the components.

Raju et al. [21] employed effective segmentation approach to categorize the kidney tumor and cysts for the ultrasound pictures. The obtained results confirmed that the suggested strategy was effective in detecting the kidney tumor and cysts. Baygin et al. [22] have presented a mechanism to identify the kidney stones utilizing the CT images. In order to choose the necessary features, the effective component evaluation task was adopted and then forwarded to the kidney stone detection phase. The suggested approach attained better outcomes utilizing the validation approaches. Expert solutions revealed that the suggested approach could help the urologists to examine the manual screenings and thus minimize the human error in the identification of kidney stones. In 2021, Yildirim et al. [23] have deployed the automated kidney stone identification from CT images with the help of deep learning algorithm. The suggested approach was observed by the experts and demonstrated that the designed task was capable to identify the kidney stones accurately. This work displayed that the deep learning approaches could be utilized to study the other difficult issues in the urology.

Ma et al. [24] have presented a strategy for the diagnosis, identification, and segmentation of the failure of chronic renal. The proposed technique was applied to the ultrasound image for pre-processing and segmentation. In the segmentation of the kidney, the recommended task was accomplished with better accuracy minimal the processing time. In 2023, Yan and Razmjooy [25] suggested the CT image-aided detection of kidney stone approach. The mechanism employed the integration of the meta-heuristics and deep learning approach. The endorsed approach produced a reliable and effective kidney stone detection method. The suggested work attained better specificity contrasted to the other validated mechanisms.

To address the drawbacks of conventional approaches, such as their prolonged computing time, and diminished accuracy, Patel and Yadav [26] examined Artificial intelligence (AI) -based algorithms for kidney tumor segmentation and prognostication. The AI methods, such as deep learning and machine learning, greatly improve the accuracy of diagnoses and make early detection easier, leading to better predictions of renal disease. Further, to segment kidney tumors and classify them, Patel et al. [27] presented a deep learning model that optimizes 3D-TR-DUnet++ for segmentation and AA-RD-GRU for classification using the Modified Crayfish Optimization Algorithm (MCOA). The model outperforms traditional methods like CNN and Residual Densenet while drastically improves accuracy and minimizes computational loads. To support physicians in quickly and accurately diagnosing benign and malignant cancers, the model enhances diagnostic accuracy, decreases computation time, and strengthens early tumor detection.

1.2 Research gaps and challenges

A kidney tumor is one of the disorder that have troubled the society. The detection of kidney tumors in the starting stage offers effective merits such as minimizing the death rates and also minimizing the side effects. Multiple approaches have been offered to detect the kidney tumors and several of the technique’s advantages and demerits are given in Table 1. In order to solve these difficulties an intelligent approach has been developed to identify the kidney tumor effectively by applying the deep learning mechanisms.

Table 1. Features and challenges of the conventional kidney tumor detection mechanism

The major contributions of the recommended kidney tumor detection approach is as follows.

•To design the kidney tumor detection framework that effectively solves the complexities of the conventional approaches and helps to detect the kidney tumors in early stages.

•To segment the pre-processed medical images employed the 3D-ATR mechanism an improvement of the conventional transformer and ResUnet. The parameters presented in this network are optimally selected by the implemented EPU-CSO.

•To design the EPU-CSO utilized the traditional CSO algorithm that assists in parameter optimization and enriches the system's effectiveness.

•To detect the kidney tumor the MT-RAN framework is adopted where the multi-scale transformer and RAN mechanisms are integrated to provide the better-classified outcomes.

•To prove the suggested kidney tumor detection approach's efficacy employed various conventional classifiers, algorithms, and segmentation metrices with divergent functionality metrics.

The suggested kidney tumor detection framework contents are organised as follows. The adaptive segmentation and classification for detecting the kidney tumor disorder are explained in Section II. In addition, Section III offers the EPU-CSO and adaptive transformer-based ResUnet model. Section IV illustrates the identification of the kidney tumor disease utilizing multi-scale transformer-based RAN. Furthermore, Section V demonstrates the outcomes of the suggested kidney tumor detection approach. Lastly, Section VI concludes the recommended kidney tumor detection approach.

3.1 Parameter optimization using EPU-CSO

The classical CSO is adopted from the circle’s geometrical features. One of the well-known object is a circle that has multiple features especially the centre, tangent line, diameter, and the perimeter. The circle is the curve which is enhanced with the same amount of distance among the middle point and the entire points. The CSO is a robust and simple approach. It produces effective solutions. In order to enhance the accuracy rate EPU-CSO approach is implemented because the conventional CSO takes the uniform random number $e$ in between 0 and 1. This may reduce the correctness of the approach. So, the new EPU-CSO approach estimates the variable $e$ according to the fitness values of the CSO. The calculation of the uniform random variable $e$ is derived in Eq. (3).

$e=\frac{b e f}{w o f}$          (3)

where, bef represents the best fitness found by the conventional CSO algorithm. This is the highest or most optimal solution encountered during the optimization process. Further, wof: Represents the worst fitness in the conventional CSO algorithm. This is the least optimal solution observed in the process.

Significance of the Ratio e in EPU-CSO

•Guiding Parameter Estimation: The ratio e can be used to estimate parameters or adjustments within the EPU-CSO algorithm.

•Range of Search Space Adjustment: If e is close to 1, it indicates that the fitness values are similar, suggesting a potentially narrower range of search space or less need for exploration. Conversely, if e is much less than 1, it indicates a larger spread between the best and worst solutions, which might necessitate a broader search space or more extensive exploration.

•Scaling of Random Variables: The value of e can influence how random variables are scaled. For example, a larger ratio might result in scaling that favors exploration of diverse areas, while a smaller ratio could focus on refining solutions within a smaller region.

•Refinement of Search Process: As the algorithm progresses, the ratio e helps adjust the search strategy. If the fitness values converge (i.e., e approaches 1), the algorithm might shift towards a more refined search or exploitation of promising regions. Conversely, if there is a wide spread, the algorithm may emphasize exploration to find better solutions.

The primary phases of the traditional CSO [30] are as follows. The starting phase of the CSO is the initialization of the attributes. In this phase, the entire dimension of each search member should be equally selected randomly.

The search member's initialization is conducted between the lower ll and the upper ul limit factors of the exploring region. This is given by Eq. (4).

$T_s=l l+p \times(u l-l l)$          (4)

In the above expression, the variable p specifies the arbitrary vector that has a range from 0 to 1.

The next phase is started when the search member position’s updated. Depending on the validated best position T_e, the place of the search member T_s is upgraded. This phase is derived in Eq. (5).

$T_s=T_e+\left(T_e-T_s\right) \times \tan (\theta)$          (5)

The importance of the angle $\theta$ in the above expression is very essential for the conventional CSO scheme’s phases of exploration and the exploitation. These are evaluated from Eq. (6) and Eq. (9).

$\mathrm{Q}(s, a)=\mathbb{E} s^{\prime}\left[R(s, a)+\gamma \max a^{\prime} Q^{\left(s^{\prime}, a^{\prime}\right)}\right]$          (6)

$\beta=\beta \times r m-\beta$          (7)

$c=\pi-\pi \times\left(\frac{I}{M I}\right)^2$          (8)

$f=1-0.9 \times\left(\frac{I}{M I}\right)^{0.5}$          (9)

The arbitrary measure is pointed as $r m$ with the range [0, 1] and the iteration counted is denoted as I. The conventional CSO utilized a constant variable as  with a limit [0, 1] which may reduce the accuracy rates. So that in the presented EPU-CSO, the variable  is estimated utilizing Eq. (3). And the maximum number of iterations is specified as $M$. The term $\beta$ in Eq. (7) changes from  to 0 when the iteration number gets increased in the process. In addition, the factor c in Eq. (8) changes from $-\pi$ to 0. Further, the measure f in Eq. (9) changes from 1 to 0. In the end, the variable  changes from $-\pi$ to 0.

The classical CSA adopts two phases to perform the executions and that is explained as follows.

Case (i) $I>(e . M I)$: This condition is enabled when the angle is $\theta=\beta \times r m$ and it can be provided to increase the exploration phase of the CSO.

Case (ii) $I<(e . M I)$: This condition generates when the overall angle time is $\theta=\beta \times r$ and that can employ to increase the phase of the exploitation.

Algorithm 1 specifies the pseudo-code of the suggested EPU-CSO algorithm and the Figure 3 offers the flowchart of the presented EPU-CSO.

EPU-CSO combines a parameter ratio e to modify the search strategy, therefore affecting the exploration and circular search strategies. This suggested approach uses the ratio e, which influences parameter estimation, search space range, and scaling of random variables, so having a more dynamic adjusting mechanism. It also adjusts the balance dynamically depending on the fitness values noted during the optimization process using the ratio $e$.

3.2 Trans-ResUnet for segmentation

The image segmentation is conducted by changing an image into a set of pixels which are indicated by the image designated. In the suggested kidney tumor detection approach, the pre-processed picture $G_a^{p p}$ is provided to the segmentation phase which is performed with the support of Trans-ResUnet. This network is the composition of the "Transformer and ResUnet". The transformer [31] approach is managing the problem of deep learning. In this task, the tokenization process is utilized by changing the provided input into a demolished 2D specks collection $\left\{G_a^j \in R^{a^2 . D} \mid j=1,2, \ldots J\right\}$, here the variable $a \times a$ denoted as the patch size, and the variable $M=$ $\frac{J W}{a^2}$ indicates the image specks amount. Further, the patch embedding is conducted. Here, the patches which are vectorized $G_a$ are linked together with the region of D dimensional embedding by adopting the understandable sequence projection. Eq. (10) formulates the positional information.

$G_0=\left[G_a^1 F ; G_a^2 F ; \ldots ; G_a^J F\right]+F_{p a s}$          (10)

The variable $F \in R^{\left(a^2 . D\right) \times E}$ specifies the "patch embedding projection" and the "position embedding" is denoted as $F_{p o s} \in$ $R^{M \times E}$.

There are V layers of “Multihead Self-Attention (MSA)” and “Multi-Layer Perceptron (MLP)” in the transformer encoder. Hence, the answer to the layer is expressed by Eq. (11) and Eq. (12):

$G_v^{\prime}=M S A\left(L N\left(G_{v-1}\right)\right)+G_{v-1}$          (11)

$G_v=M L P\left(L N\left(G_v^{\prime}\right)\right)+G_v^{\prime}$          (12)

The attribute LN points to the normalization of the layer and the variable $G_v$ terms the encoded representation of the image.

Figure 4 shows the Trans-Resnet architecture for the segmentation that includes encoding, decoding and attention mechanism. Where the encoding path narrows down the spatial dimensions while extracting features, and the decoding path up samples the feature maps to reconstruct the original image with segmentation information. The attention mechanism helps the model focus on relevant parts of the image, improving its performance.

Because of the attribute lost identities in the deep NN implemented by the minimized weight attribute gradients maximizes the degradation. The ResUnet [32] is the deep U-net that is an encoder-decoder framework for the semantic segmentation. The ResUnet combines the three paths of coding into the framework of encoding-decoding. The ResUnet utilizes the features of both deep residual learning and Unet. The ResUnet framework is the fully CNN with minimal attributes that focuses to attain better functionality. It is the improved version of the conventional U-Net framework. The ResUnet includes a decoding and encoding network and to connect these two networks there is a bridge. The U-net utilizes two 3×3 convolutions. Every convolution has an activation function. When focusing on the ResUnet, these layers are exchanged by the residual block. The TransResUnet framework’s architecture is illustrated.

4.png

Figure 4. The architecture of the Trans-ResUnet for the image segmentation in the suggested kidney tumor detection approach

3.3 Suggested 3D-AT-ResUnet for segmentation

The Trans-ResUnet is a very compatible and robust approach. This approach is very efficient for the medical image segmentation. However, the network has high computational requirements and utilizes more amounts of parameters. So, the suggested kidney tumor detection mechanism presented the 3D-ATR mechanism for the effective segmentation approach. Here, the pre-processed image $G_a^{p p}$ is subjected as an input. This new mechanism executes the images in an “end-to-end” way and performs the 3D evaluation on the both decoder and encoder sides. This approach employs the 3D convolution for the effective functionality and successive outcomes. The 3D-ATR approach overcomes the complexities of the conventional approaches. The transformer-ResUnet includes several attributes such as hidden neurons, step per epoch, and the number of epochs. The high amount of these attributes raises the network complexities and leads to overfitting. So, that these attributes are optimally tuned in this suggested mechanisms. For that purpose, the developed EPU-CSO is utilized. The objective function of this task is given by Eq. (13).

$o b_1=\underset{\left\{hn^{3 D-A T R}, e p^{3 D-A T R}, s e^{3 D-A T R}\right\}}{\arg \max }[D c+A c]$          (13)

Here, the steps per epochs in 3D-AT-ResUnet are given as $s e^{3 D-A T R}$ and the hidden neuron counts in 3D-AT-ResUnet are specified as  $h n^{3 D-A T R}$, and the number of epochs in 3D-AT-ResUnet is denoted as $e p^{3 D-A T R}$. Moreover, the steps per epoch in 3D-AT-ResUnet vary from 300 to 1000 and the hidden neuron counts in 3D-AT-ResUnet are changed from 5 to 255. Also, the “number of epochs” in 3D-AT-ResUnet falls from 5 to 50. The variable $D c$ indicates the “dice coefficient” which is the overlap among the segmented and masked pictures. Besides, the variable $A c$ specifies the accuracy which finds the relationship among the original and processed images. The formulations of these two variables are given in Eq. (14) and Eq. (15):

$D c\left(G_a, G_a^{ {seg }}\right)=\frac{2\left(G_a \cap G_a^{\text {seg }}\right)}{G_a+G_a^{ {seg }}}$          (14)

$A c=\frac{s a+n t}{s a+n t+h i+y a}$          (15)

where, the variable $G_a^{s e g}$ indicates the segmented image, and the factor ya point to the “false negative”. The term hi denotes the “false positive” and the variable nt refers to the “true negative”. Also, the factor sa indicates the “true positive”. In the end, the effective segmented images $G_a^{s e g}$ are attained from the suggested 3D-AT-ResUnet framework. The 3D-AT-ResUnet appears in Figure 5, showcasing parameter tuning. The visual displays the 3D AT-Res U-Net architecture, which is a specialized deep learning model created for the purpose of segmenting medical images. This model is an enhanced version of the conventional U-Net design, which integrates attention mechanisms and residual connections to enhance performance.

5.png

Figure 5. The framework of the 3D-AT-ResUnet mechanism for the segmentation approach with parameter optimization

5.1 Simulation setup

Python platform has been used to process the recommended kidney tumor detection mechanism and remarkable outcomes were achieved. The proposed kidney tumour detection approach employs 10 different combinations of the parameters at each generation i.e., 10 populations and each solution in these populations is characterized by 3 parameters or 3 chromosome lengths for segmentation, classification and overall architecture. Moreover, the maximum iteration count of the recommended kidney tumor detection framework was 50. Some of the traditional optimization algorithms like Tuna Swarm Optimization (TSO) [35], Coyote Optimization Algorithm (COA) [36], Fire Hawk Optimizer (FHO) [37], and CSO [28] are utilized for the evaluation process. In addition, several segmentations approach like Unet [38], ResUnet [32], ResUnet++[39], and 3D-ATR [40] were adopted. Also, the detection mechanisms such as CNN [41], LSTM [42], Mobilenet [43], and Resnet [44] were employed for the estimation of the recommended kidney tumor detection mechanism.

5.2 Evaluation metrics

The suggested kidney tumor detection mechanism utilized several performance metrics for the evaluation. These are given as follows.

Dice coefficient: Measures the similarity between two sets, typically the predicted and ground truth segmentations. It ranges from 0 (no overlap) to 1 (perfect overlap). It is formulated in Eq. (14). In medical imaging, the Dice coefficient determines how closely the ground truth region identified by an expert overlaps the expected tumor region.

Accuracy: The proportion of correctly classified pixels (or regions) out of all the pixels. It is formulated in Eq. (15).

In image classification or segmentation, it indicates the algorithm's accuracy in accurately identifying both positive (e.g., tumor) and negative (e.g., non-tumor) pixels.

Jaccard coefficient: It is employed to validate the percentage of the overlap among the gatherings.

$J c\left(G_a, G_a^{s e g}\right)=\frac{\left|G_a \cap G_a^{s e g}\right|}{\left|G_a \cup G_a^{s e g}\right|}$          (28)

where, G_a represents original image, and $G_a^{s e g}$ is segmented image. Jaccard coefficient is used to evaluate how well the algorithm detects the correct area of an object like a tumor by computing the overlap between the predicted and actual regions.

Precision: It is the description of the arbitrary errors and estimates the variability of the statistics.

$P=\frac{s a}{s a+n t}$          (29)

It measures how accurate the model is in predicting true positives without falsely predicting non-tumor regions as tumors.

NPV: It estimates the rate of true negative estimations concerning overall negative estimations.

$N P V=\frac{h i}{h i+y a}$          (30)

NPV in medical imaging refers to the probability that areas identified as non-tumor are indeed non-tumor, hence offering confidence in the negatives.

FPR: It is the factor to estimate the probability of the correctness.

$F P R=\frac{n t}{n t+h i}$          (31)

FPR analyzes the rate at which the model mistakenly labels non-tumor areas as tumors, therefore causing unwarranted treatments.

F1-score: It integrates the scores of the precision and recall.

$F 1-Score =\frac{2 \times s a}{2 \times s a+n t+y a}$          (32)

The F1-score gives an overall sense of the balance between precision (accuracy of positive predictions) and recall (sensitivity).

FDR: It is utilized to correct the various comparisons.

$F D R=\frac{n t}{n t+s d}$          (33)

FDR assesses how often the model incorrectly predicts a positive (e.g., a tumor) when it’s actually not, indicating potential over-diagnosis.

Sensitivity: It is the likelihood of a positive test outcome.

$Sen =\frac{s a}{s a+y d}$          (34)

Sensitivity measures how effectively the model detects all actual positive cases (e.g., all tumors), critical for ensuring that important anomalies are not missed.

Specificity: It evaluates how well the experiment detects the true negative.

$s p e c=\frac{h i}{h i+n t}$          (35)

Specificity measures how well the model identifies non-tumor areas, important to avoid false positives that could lead to unnecessary treatments.

MCC: It is the correlation factor among the monitored and estimated classifications.

$M C C=\frac{s a \times h i-n t \times y a}{\sqrt{(y a+s a)(y a+h i)(s a+n t)(h i+n t)}}$          (36)

MCC is useful for binary classification problems (e.g., tumor vs. non-tumor) and is a more reliable metric when class sizes are imbalanced. For all above evaluation metrics, we have used the folowing terms for notation: G_a – original image, $G_s^{\arg }$ – segmented image, sa – true positives, nt- false positives, ya – false negatives, and hi – true positives.

5.3 Confusion matrix estimation of the recommended kidney tumor detection approach

The recommended kidney tumor detection mechanism is processed with the confusion matrix to estimate the accuracy of the network for the two datasets. This is presented in Figure 9. The accuracy of the suggested kidney tumor detection approach is enriched by 96.18% when considering the second data resource in Figure 9(b) (The presented accuracy scores for Dataset 1 and Dataset 2, is 95.38% and 96.18%, respectively, provides additional evidence that supports this observation). From this, it is highlighted that the designed kidney tumor identification mechanism has better efficacy.

9a.png

9b.png

Figure 9. The confusion matrix evaluation of the recommended kidney tumor detection framework concerning, (a) Dataset 1, and (b) Dataset 2

5.4 ROC investigation of the suggested kidney tumor detection mechanism over various classifiers

Figure 10 shows the ROC investigation of the developed kidney tumor detection framework against various classifiers for the two data resources. The ROC is validated based on the “true and false” positive rates. For the first data resource in Figure 10(a), the ROC of the suggested kidney tumor mechanism is advanced by 45% of CNN, 46% of LSTM, 50% of Mobilenet, and 53% of Resnet accordingly when taking the “false positive rate” value is 0.6. Hence, the suggested kidney tumor detection approach has outperformed other classifiers.

10a.png

10b.png

Figure 10. The ROC calculation of the suggested kidney tumor identification framework over various conventional classifiers concerning, (a) Dataset 1, and (b) Dataset 2

5.5 Convergence estimation of the designed EPU-CSO algorithm over diverse conventional algorithms

The designed EPU-CSO algorithm is validated with the convergence measure against diverse traditional algorithms and provided in Figure 11 for the two data sources. The amount of iteration is employed for this process. For the second data source in Figure 11 (b), the convergence of the suggested EPU-CSO scheme has maximized by 91% of TSO, 91% of COA, 92% of FHO, and 92.3% of CSO appropriately when the number of iterations is 30. The outcomes highlight that the recommended EPU-CSO has better convergence rates.

11a.png

11b.png

Figure 11. The convergence investigation of the suggested EPU-CSO algorithm over diverse optimization algorithms regarding, (a) Dataset 1, and (b) Dataset 2

5.6 Statistical report of the improved EPU-CSO algorithm over multiple optimization algorithms

Table 2 offers the surprising statistical outcomes of the designed EPU-CSO scheme against various conventional algorithms for the two datasets. The suggested EPU-CSO has improved by 30% of TSO, 16% of COA, 14% of FHO, and 14.2% of CSO correspondingly when considering the best factor in the first dataset. From this, it is confirmed that the developed EPU-CSO has higher functionalities. Looking at best, mean, median values, EPU-CSO shows the best performance on both datasets; this indicates that it is often able to identify better solutions on average and also performs regularly well. Although CSO is not the best performer in all measures, typically it shows a balance of strong performance with low variability in findings on Dataset-1.

Table 2. The statistical validation of the designed EPU-CSO algorithm over various optimization algorithms for two datasets

5.7 Functionality validation of the suggested segmentation approach over diverse algorithms and segmentation methods

The functionality of the recommended 3D-AT-ResUnet-assisted segmentation task is estimated for the two datasets against various algorithms and segmentation approaches and demonstrated in Figure 12 and Figure 13. When considering the dice coefficient for the first data source in Figure 13 (b), the segmentation process of the recommended kidney tumor detection has enhanced by 85% of Unet, 88% of ResUnet, 91% of ResUnet++, and 93% of 3D-ATR correspondingly. This shows that the suggested 3D-AT-ResUnet-assisted segmentation task in developed kidney tumor detection has effective solutions.

12a.png

12b.png

12c.png

Figure 12. The functionality validation of the designed segmentation process in the kidney tumor detection framework over various optimization algorithms regarding, (a) Accuracy, (b) Dice coefficient, and (c) Jaccard

13a.png

13b.png

13c.png

Figure 13. The functionality estimation of the recommended segmentation process in the kidney tumor detection framework over various segmentation approaches regarding, (a) Accuracy, (b) Dice coefficient, and (c) Jaccard

5.8 Comparative estimation of the recommended kidney tumor detection mechanism over multiple algorithms and classifiers

The comparative evaluation of the recommended kidney tumor detection framework over distinct algorithms and classifiers is presented in Figures 14-17 for the two datasets. Several evaluation metrics are utilized to estimate the functionality of the approach. From Figure 14(a), the specificity of the suggested kidney tumor detection approach is enriched by 85% of TSO, 85.5% of COA, 86% of FHO, and 86.5% of CSO appropriately for the first data source over FPR. Thus, the recommended kidney tumor detection task has better functionalities against the pre-existing methodologies.

14a.png

14b.png

14c.png

Figure 14. The comparative research of the designed kidney tumor detection framework over various conventional algorithms for the first data source concerning (a) Sensitivity Vs FPR, (b) Precision Vs FDR, and (c) NPV Vs FOR

15a.png

15b.png

15c.png

Figure 15. The comparative research of the suggested kidney tumor detection framework over various conventional classifiers for the first data source concerning (a) Sensitivity Vs FPR, (b) Precision Vs FDR, and (c) NPV Vs FOR

16a.png

16b.png

16c.png

Figure 16. The comparative research of the suggested kidney tumor detection framework over various conventional algorithms for the second data source concerning (a) Sensitivity Vs FPR, (b) Precision Vs FDR, and (c) NPV Vs FOR

17a.png

17b.png

17c.png

Figure 17. The comparative research of the suggested kidney tumor detection framework over various conventional classifiers for the second data source concerning (a) Sensitivity Vs FPR, (b) Precision Vs FDR, and (c) NPV Vs FOR

5.9 Performance evaluation of the suggested kidney tumor detection approach over various algorithms and classifiers

Figures 18-21 illustrate the performance investigation of the recommended kidney failure detection strategy against distinct algorithms and classifiers for the two data resources. By considering the epoch values and some of the performance factors, this operation is performed. For the second data source in Figure 20(a), when the epoch value is 300, the accuracy of the suggested kidney tumor detection mechanism is maximized by 70.8% of CNN, 70.4% of LSTM, 69.3% of Mobilenet and 68.9% of Resnet appropriately. Thus, the suggested kidney tumor detection approach has enriched performance rates against other classical methodologies.

18a.png

18b.png

18c.png

Figure 18. The performance calculation of the suggested kidney tumor identification framework over various conventional algorithms for the first data source concerning, (a) Accuracy, (b) FNR, (c) FPR, (d) NPV, (e) Precision, (f) Sensitivity, and (g) Specificity

19a.png

19b.png

19c.png

Figure 19. The performance calculation of the suggested kidney tumor detection framework over various conventional classifiers for the first data source concerning, (a) Accuracy, (b) FNR, (c) FPR, (d) NPV, (e) Precision, (f) Sensitivity, and (g) Specificity

20a.png

20b.png

20c.png

Figure 20. The performance calculation of the suggested kidney tumor detection framework over various conventional algorithms for the second data source concerning, (a) Accuracy, (b) FNR, (c) FPR, (d) NPV, (e) Precision, (f) Sensitivity, and (c) Specificity

21a.png

21b.png

21c.png

Figure 21. The performance calculation of the suggested kidney tumor detection framework over various conventional classifiers for the second data source concerning, (a) Accuracy, (b) FNR, (c) FPR, (d) NPV, (e) Precision, (f) Sensitivity, and (c) Specificity

5.10 Overall comparative evaluation of the recommended kidney tumor detection approach over various conventional algorithms and classifiers

Table 3 and Table 4 offer the overall comparative validation of the recommended kidney tumor detection mechanism over various algorithms and classifiers for two datasets. The specificity of the recommended kidney tumor detection approach is enhanced by 7.9% of TSO, 6% of COA, 4.1% of FHO, and 2.4% of CSO correspondingly when taking the first dataset in Table 3. Hence, the outcomes explained that the suggested kidney tumor detection mechanism has better functionalities.

Table 3. The overall comparative estimation of the designed kidney tumor identification strategy over diverse traditional algorithms for two datasets

Table 4. The overall comparative estimation of the designed kidney tumor identification strategy over diverse traditional classifiers for two datasets