A related convolutional neural network for cancer diagnosis using microRNA data classification Authors: Najmeh Sadat Jaddi, Salwani Abdullah, Say Leng Goh, Mohammad Kamrul Hasan Published: 2024 ABSTRACT AbstractThis paper develops a method for cancer classification from microRNA data using a convolutional neural network (CNN)‐based model optimized by genetic algorithm. The convolutional neural network has performed well in various recognition and perception tasks. This paper contributes to the cancer classification using a union of two CNNs. The method's performance is boosted by the relationship between CNNs and exchanging knowledge between them. Besides, communication between small sizes of CNNs reduces the need for large size CNNs and, consequently, the computational time and memory usage while preserving high accuracy. The method proposed is tested on microRNA dataset containing the genomic information of 8129 patients for 29 different types of cancer with 1046 gene expression. The classification accuracy of the selected genes obtained by the proposed approach is compared with the accuracy of 22 well‐known classifiers on a real‐world dataset. The classification accuracy of each cancer type is also ranked with the results of 77 classifiers reported in previous works. The proposed approach shows accuracy of 100% in 24 out of 29 classes and in seven cases out of 29, the method achieved 100% accuracy that no classifier in other studies has reached. Performance analysis is performed using performance metrics.Keywords: biology computing, cancer, convolutional neural nets INTRODUCTION A large illustration and superior vision of many genetic changes related to cancer is presented by parallel quantification of thousands of gene expressions. The word cancer is employed to recognize a disease where there is abnormal cell division with no control. This causes a lump called a tumour which expands to other parts of the body via the lymph systems and blood vessels. A range of cancer types and more than100 different tumours are often named by their positioning in the body [1]. An accurate and early prognosis of cancer assists in a correct and good line of treatment. Since the existing patterns of microRNA can be associated with cancer type, microRNA classification can be employed as an instrument for cancer prediction and diagnosis. The data provided by microarray technology have shown huge potential in cancer classification and diagnosis [2]. The microarray technology produces the datasets normally with thousands of gene expressions gained from each genetic sample. The complexity and high dimension of cancer data obtained by microarray cause difficulty in its analysis, classification, and prognosis. Different methods have been used to classify cancer datasets in previous works [3, 4, 5, 6, 7, 8]. Neural networks have been used by many researchers in the literature [9, 10, 11, 12, 13]. Deep neural networks have been employed in a variety of applications in literature. Less number of CNNs has been used for non‐image data [14, 15, 16, 17, 18] compared to the huge number of CNNs applied to image processing [19, 20]. Among all the research in machine learning, much research for cancer diagnosis problems has recently been done. The success of researchers in [3–6, 21, 22] using machine learning for cancer diagnosis is a motivation for using strong and boosted deep learning for this problem. In these researches machine learning and mostly deep learning using CNNs are offered and their results show successful improvement in cancer diagnosis. Although it is proven that setting up more layers and more parameters for CNN often develops a model with higher accuracy, big models are too expensive in computational complexity and memory cost. However, large networks have the drawback of high computational complexity and memory cost. One approach to decrease the computational time and memory usage is somehow reducing the number of parameters in the network. Since the convolution layers of CNN cause a lot of repeated calculations, this computation relies on high performance with additional time and memory. Some methods offered in past research to speed up convolutions while preserving the performance of the network are: sparse net [23], bottleneck layers [24], wider convolutions [25], and depth‐wise separable convolutions [26]. Model compression is another method for speeding up CNN process. A survey of model compression and acceleration for CNN is presented in [27]. To surmount the above difficulty, in this paper, we propose a relational convolutional neural network (R‐ConvNet) which involves two smaller CNNs, each being active within a certain number of iterations. In this case, once one of the CNNs is active the knowledge from the previously active CNN is used to aid the progress in the CNN process. After a certain number of iterations, this currently active CNN becomes inactive and its knowledge is transferred to the CNN becoming active. In words, the knowledge is exchanged between the active/inactive CNNs. The knowledge of the last feature map of the currently active CNN is transferred to the future active CNN in the last feature map. Therefore, the active CNN has the knowledge from the training process besides the knowledge transferred from the previously active CNN. In this method, the need for the large size source model is eliminated and two CNNs are boosting each other. A genetic algorithm (GA) is employed to optimize the parameters of the CNNs. A combination of single‐point and uniform crossovers is used and random mutation is employed to generate new chromosomes during the search process R‐ConvNet is tested on microRNA data. The rest of this paper is organized as follows: Related works are given in Section 2. Section 3 provides the details of the proposed method. Section 4 discusses the results of experiments to assess the performance of the R‐ConvNet and Section 5concludes the performance of the R‐ConvNet and the results obtained. RELATED WORKS One of the methods used in past research to decrease the effect of complex computation in convolution layers is sparse net. Sparse net was first proposed in [23] to reduce the computation time and effort. Later, an accelerator for compressed‐sparse for CNN was proposed in [28]. Another research [29] proposed an Interleaved structured sparse CNN. An accelerator for sparse CNN on FPGAS was offered in [30]. A novel CNN‐based fault recognition method via image fusion of multi‐vibration‐signals was presented in [31] and wide network with bottlenecks was proposed in [32]. Some recently proposed are available in [33, 34, 35, 36, 37]. An additional way to accelerate the CNNs available in the literature is network compression. Compression of convolutional neural networks on mobile devices is proposed in [38]. Merging similar neurons for deep network compression was recently proposed in [39]. A lossless compression of deep networks was published in [40]. Knowledge distillation (KD) is the most commonly used for network compression in the literature [41, 42, 43, 44, 45, 46, 47, 48, 49]. KD methods commonly train a small model while the knowledge from a large model is transferred to the small model. The large model is called the source/teacher model and the small model is called the target/student model. Some applications of model compression on object detection [42], mobile devices [50, 51], visual relationship detection [52], and character recognition [53, 54] are available in the literature. Another method to reduce the complexity of computation in deep networks is to use the boosted small size of the networks to achieve the accuracy as high as a CNN with a large size architecture and expensive time and memory usage. Boosting the accuracy of multispectral image pan sharpening by learning a deep residual network was proposed in [55]. Incremental boosting CNN for facial action unit recognition was studied in [56]. Another boosting method was offered for convolutional neural networks [57]. A boosted CNN for object recognition at a large scale was suggested in [58]. Simulated perfusion MRI data to boost training of CNNs for lesion fate prediction in acute stroke was presented in [59]. PROPOSED METHOD The proposed method, R‐ConvNet involves communication between two CNNs in which each one sends its knowledge from the last feature map to another active CNN in a pre‐determined number of iterations. The optimization of the model is performed using GA. The crossover and mutation operators of the GA are used to form the children. The details of the R‐ConvNet are presented in the following sub‐sections. Relationship between CNNs In R‐ConvNet two CNNs are involved, in which one of them is active at anytime. Cooperation is performed between CNNs to boost each other by exchanging knowledge. This knowledge exchange is carried out from the last feature map of the previously active CNN into the currently active one. It means that the transferred feature map and the last feature map of the currently training CNN is involved in the fully connected layer of CNN. The process of cooperation between CNNs is as follows: If we consider CNN1 is active and involved in the training process first and CNN2 is inactive, CNN1 turns inactive after a certain number of iterations and CNN2 (as an active CNN) starts the training process and employs the knowledge from CNN1 in which the last feature map from the CNN1 and the last feature map from the CNN2 is involved in the fully connected layer of CNN2. After a pre‐determined number of iterations, CNN2 becomes inactive and, consequently, CNN1 becomes active and follows the training process. During the training process, CNN1 uses the knowledge from CNN2. This flip‐flop cooperation process is continued until a termination criterion is reached. Figure 1 shows an example of the relationship of the two CNNs during the training process. This figure, with the aid of an example, clarifies the flip‐flop cooperation between CNNs. This figure starts from raw data going through the kernel to fill up the convolutional layer, max pooling and fully connected layer consecutively. As shown in this figure, not only the output of the last feature map from the active CNN, but also the output of the last feature map from inactive CNN is used as input for a fully connected layer. In this case, the last feature map from the currently active CNN will be used for the next active CNN. In other words, to transfer knowledge between two CNNs, the values of the last feature map from inactive CNN are fed to the fully connected layer of active CNN as input along with its feature map values. The fitness function which is the accuracy of the classification is calculated for the new model. The cooperation between CNNs in this manner provides exchanging knowledge several times and boosting each other using the knowledge supplied by the partner CNN during several previous iterations. This act somehow re‐starts the training process from a possibly higher‐quality point many times and, consequently, results in improvement each time. This guides the algorithm towards an optimum solution more efficiently. The pseudocode of this relationship is provided in Algorithm 1. An example of relationship process of CNNs during the training process. Pseudocode of cooperation between CNNs. while (stopping criteria is reached) do the training process for active CNN update the best model if pre‐determined number of iterations is achieved change the active CNN transfer knowledge from inactive into active CNN endif increase the iteration endwhile return the best model The purpose of using this cooperation is to employ CNN with high accuracy while there is no need to train a very large network with the drawback of high computational time and memory cost. In this method, even using CNNs with small sizes can achieve a reasonable performance since each CNN tries to share its knowledge with the other one. This cooperation results in better convergence towards the optimum solution with less computational time and memory usage. Learning process and optimization by genetic algorithm During this study, after data collection and preparation, the GA is used to train the R‐ConvNet. The training and test procedure and communication between GA and R‐ConvNet required a technical process to perform high‐quality training and find the best CNN model. In this optimization process, training of CNN is performed using a training set of data. In R‐ConvNet the cooperation of two CNNs is carried out to perform a fast and lesser computational time process while preserving the high accuracy of the prediction. In the training stage of the algorithm, an initial population with randomly encoded chromosomes is generated. Selection of parents is performed and crossover and mutation are carried out to form children. If CNN1 is active, child1 is used to generate a new model while if CNN2 is active, child2 is employed to generate the new model. The relationship between two CNNs is done and the knowledge transmission is performed in a pre‐determined number of iterations to improve the performance of this algorithm. The fitness function of the candidate model is evaluated using the training set and the best model is updated. This process is continued until a termination criterion is reached. The best (optimal) model found by the training stage is sent to the testing stage, the prediction is executed using testing data and the classification accuracy is calculated. The step‐by‐step learning process is given as follows: Step 1: Initialization Step 1: Initialization The initial population is randomly set. Each solution in the population is a vector containing the trainable network's parameters. Next, they are integrated into an array representing a chromosome of the nets. Therefore each solution (chromosome) shows a network's model which is demonstrated as Y i (i = 1, 2,…, PopSize). A fully folded chromosome is employed in which each element of this chromosome holds either a convolution layer as a whole or a fully connected layer. Each element of a convolution layer holds all values of kernel parameters and each element of a fully connected layer holds all nodes in which each element of a node holds connection weight values. This chromosome representation benefits faster and easier GA operations. Step 2: Calculating objective function and updating the best model Step 2: Calculating objective function and updating the best model Each solution randomly located in the search space is evaluated by calculating the accuracy of the classification of the candidate network's architectures which is shown as y = {e1(y), e1(y), …., eq(y)} where, e denotes the objective function of e→B→S from the set of real numbers, y describes the problem variable and q represents the count of objectives. The best model with higher accuracy of the classification is updated. Step 3: Parent selection Step 3: Parent selection Population size is shown by PopSize, and the fitness value of individual i is expressed by fi. The probability of the individual being selected Pi is calculated by Equation (1). The fitness value fi for each individual in the population is calculated. Step 4: Generation Step 4: Generation In this work, a two‐step crossover is proposed which combines a single‐point crossover with a uniform crossover. This combination aims to avoid similar children generated by single‐point using an additional step of uniform crossover. In the single‐point step, a cut‐off point on both parents is selected randomly and then that the two sections divided by that point in the parent vectors is swapped between the two parents. The uniform step generates a random binary vector. The cells in this vector decide the genes in the child's chromosome. Selection of genes is performed from the first parent where the cell at the same position of the vector is 1. If the cell in the same position of the vector is 0 the gene from the second parent is selected. A combination of the genes forms a child. The second child is generated in vice versa. If a cell in binary vector is 1 the gene from the second parent is chosen while if it is 0 the gene from the first parent is selected. To form final children, a random alteration of a gene as a mutation operation is performed. The probability of crossover and mutation are given in Table 2. Step 5: Child selection Step 5: Child selection TABLE 2Characteristics of R‐ConvNet parameters.ParameterDescriptionValueNumber of output featuresKernel sizePopSizePopulation size100––NumGenNumber of generations250––Pc Probablity of crossover0.5––Pm Probablity of mutation0.07––KtrKnowledge transmission rate10––InInput layer–1–C1Convolution layer1–31×7P1Pooling layer1–31×3C2Convolution layer 2–91×5P2Pooling layer2–91×3FC1Fully connected layer1–98–FC2Fully connected layer2–64– Characteristics of R‐ConvNet parameters. If CNN1 is active, child1 is used to generate a new model and if CNN2 is active, child2 is fed to CNN2 and the new model is generated. Step 6: Convolutional layer Step 6: Convolutional layer A convolutional layer takes n input feature maps and produces m feature maps as output. A convolution action is performed for each input feature map using a shifting window with a shifting stride of s = 1 and kernel size of k × k (in this study1×3). Then summation of the individual convolutional results is carried out. A value of bias is added to each summation output. We consider the simple case when the stride is 1 and no padding is used. Hence, we have y (or x i+1) in R Hi+1×Wi+1 , with Hi+1 = Hi − H + 1, and Wi+1 = Wi − W + 1. Equation (2) shows output feature maps in a convolutional layer mathematically, Step 7: Pooling layer Step 7: Pooling layer To reduce the feature dimensions, pooling layers in CNNs are used. The pooling layer summarizes the outputs of previous neurons in the feature map. Max pooling is used to generate a feature map. Equation (3) shows the pooling process mathematically where P stands for a maximum selection of neighbouring elements and s represents the stride: Step 8: Fully connected layer Step 8: Fully connected layer All nodes from the input layer to the output layer are fully connected in fully connected layers. The output vector in a fully connected layer is mathematically calculated as in Equation (4) and its generalized formula is shown in Equation (5). Equations (4) and (5) are used for when basically a CNN is active and is training but at the knowledge transmission stage the Equation (6) and its generalized Equation (7) are used for the output vector. In these formulas LFP denoted as last feature map. LFP1 is for last feature map of CNN1 and LFP2 is for last feature map of CNN2. The computation in fully connected layer can be seen as m group of vectors, with vector length n. Step 9: Flip‐flop cooperation between CNNs and knowledge transmission Step 9: Flip‐flop cooperation between CNNs and knowledge transmission Both the last feature map (LFP) from currently inactive CNN and currently active CNN are used to feed the fully connected neural network in active CNN, if a certain number of iterations (knowledge transmission rate, Ktr) is completed and it is time to change the active CNN. where q = certain number of iteration. Step 10: Fitness function evaluation Step 10: Fitness function evaluation The fitness function is the accuracy of the classification, which is calculated based on the new model produced by the CNNs. Step 11: Updating the best model Step 11: Updating the best model If the new model obtained a better accuracy compared to the best model found so far, then the new model is kept as the best model and its accuracy is updated as the best fitness function so far. Step 12: Termination Step 12: Termination The number of generations is the termination criterion and the search process is terminated if the number of generations assigned in the initial stage is reached. The architecture of the CNNs is not necessarily too large since the relationship between the two CNNs results in a boosted algorithm and leads the algorithm towards a better solution and therefore better accuracy of the classification. EXPERIMENTAL RESULTS In this section, the results of an experiment conducted on microRNA data are presented and compared. The characteristics of the data which contains 29 types of cancers are shown in Table 1. A total of 1046 features (genes) are observed in this dataset and the data contains information about 8136 patients. The range of a number of samples is extremely unbalanced from 778 for class BRCA to only 35 for class CHOL. TABLE 1Characteristics of 29 cancers in dataset.ClassDescriptionNumber of samplesACCAdrenocortical carcinoma80BLCABladder urothelial carcinoma415BRCABreast invasive carcinoma778CESCCervical squamous cell carcinoma and endocervical adenocarcinoma308CHOLCholangiocarcinoma35ESCAEsophageal carcinoma47FPPPFFPE pilot phase II200HNSCHead and neck squamous cell carcinoma45KICHKidney chromophobe488KIRCKidney renal clear cell carcinoma66KIRPKidney renal papillary cell carcinoma261LGGLower grade glioma292LIHCLiver hepatocellular carcinoma530LUADLung adenocarcinoma374LUSCLung squamous cell carcinoma458DLBCLymphoid neoplasm diffuse large B‐cell lymphoma341MESOMesothelioma87PAADPancreatic adenocarcinoma179PCPGPheochromocytoma and paraganglioma184PRADProstate adenocarcinoma500SARCSarcoma262SKCMSkin cutaneous melanoma452STADStomach adenocarcinoma399TGCTTesticular germ cell tumours155THYMThymoma514THCAThyroid carcinoma124UCSUterine Carcinosarcoma418UCECUterine corpus endometrial carcinoma57UVMUveal melanoma80 Characteristics of 29 cancers in dataset. Before performing any procedure, pre‐processing was carried out to prepare the data for the classification task. 53 genes contained a value of 0 for all samples. Therefore, these genes were removed and the number of genes was decreased to 993.Seven patients have incomplete records and thus these patients were also removed. Therefore, the number of patients was decreased to 8129. Normalization and discretization was performed. Pre‐analysis was conducted after normalization and the distribution of the data is found. Based on this pre‐analysis, most of the data were in the range of [0, 0.5] (in normalized data). Therefore, smaller intervals were used for the range of [0, 0.5] and bigger intervals were employed in the range of [0.5, 1] in carrying out discretization. To validate the robustness of the proposed method, tenfold cross‐validation was used. With this method, we have one data set which we divide randomly into ten parts. We use nine of those parts for training and reserve one‐tenth for testing. We repeat this procedure ten times each time reserving a different tenth for testing. To prepare the data for tenfold cross‐validation, the data were randomly split into two parts. 7320 cases for the training set and 816 cases for the testing set were considered. In this way, 90% of the data were used for the training set and 10% of the data were employed for the testing set. The training and testing stages are performed using the relationship between CNNs. The data with a higher number of training sets (90%) prevents over fitting in the training process. Another characteristic of the proposed method to prevent over fitting is less complexity of the model resulted by less number of layers and parameters when we use the two small CNNs instead of having a large CNN. In addition, early stopping was also performed by monitoring the model's performance on a validation set and stopping training when performance degrades. The experiment conducted in this research to perform the classification of microRNA data is shown in Figure 2. Overall process of R‐ConvNet to perform microRNA classification. The R‐ConvNet parameters used in this study are presented in Table 2. These parameters are the best combination found after several runs. In this study, two convolution layers were used and two pooling layers were structured between them to reduce the size of output features. As the problem domain is classification, each row of data is a map of the corresponding sample and therefore one‐dimensional shifting window was considered as the kernel for the convolution layers and the pooling layers. The convolution and the pooling work with a stride of 1. Table 3 provides the classification accuracy, targeting each class separately. This table shows that the method obtained 100% classification accuracy in all classes except in FPPP, KIRC, LUAD, PRAD, UCEC, which is only five classes out of 29 classes. As is shown in this table, the method could obtain an accuracy of 100%, even for classes with a small number of samples such as CHOL with only 35 samples, DLBC with 47 samples. UCS with 57 samples, and KICH with 66 samples. This proves the strong ability of the method to classify imbalanced data. We believe that the high ability of the method is due to firstly using CNN as a deep neural network and secondly communication between them and boosting each other. These advantages of the proposed method result in the high performance of classifiers in cancer‐type classification of high‐dimensional microRNA data. TABLE 3Classification accuracy of each class.ClassNumber of samplesClassification accuracy of R‐ConvNetACC801BLCA4151BRCA7781CESC3081CHOL351DLBC471ESCA2001FPPP450.854585HNSC4881KICH661KIRC2610.962382KIRP2921LGG5301LIHC3741LUAD4580.979341LUSC3411MESO871PAAD1791PCPG1841PRAD5000.990234SARC2621SKCM4521STAD3991TGCT1551THCA5141THYM1241UCEC4180.984629UCS571UVM801 Classification accuracy of each class. In addition, in order to have a better comparison, the range of classification accuracy of the R‐ConvNet was ranked with 77 other classifiers in the state‐of‐the‐art reported in [60]. The comparison is shown in Table 4. In this comparison, classification accuracy of each class (cancer type) is being focused on. The cells marked with * show the range of classification accuracy that the R‐ConvNet is located in. In this comparison, the methods with accuracy below 50% are not considered. Based on this table, FPPP is the only class of cancer for which the proposed approach is placed in the range of accuracy > 85%. Four of the classes (KIRC, LUAD, PRAD, and UCEC) were in range of accuracy > 90% and the rest have shown accuracy 100%. In seven cases (ACC, CESC, CHOL, DLBC, ESCA, HNSC, and MESO) out of 29, the method achieved 100% classification accuracy that no classifier in previous works can reach. TABLE 4Ranking of classification accuracy of each cancer type using R‐ConvNetagainst 77 other previous classifiers.ClassClassifiers with accuracy 100Classifiers with accuracy > 90%Classifiers with accuracy > 80%Classifiers with accuracy > 75%Classifiers with accuracy > 50%ACC1* 28384955BLCA2* 6202143BRCA9* 33434859CESC1* 3121636CHOL1* 05519DLBC1* 2112044ESCA1* 1101034FPPP003* 216HNSC1* 1111KICH2* 22354152KIRC4151* 505055KIRP2* 20444853LGG22* 49535456LIHC59* 65676769LUAD1846* 535458LUSC4* 32343746MESO1* 5121635PAAD3* 6303644PCPG51* 57616161PRAD555* 575758SARC2* 21263445SKCM6* 45545661STAD2* 23353651TGCT32* 47535358THCA3* 12363754THYM33* 58606164UCEC24* 41130UCS3* 14273449UVM2* 15354957 Ranking of classification accuracy of each cancer type using R‐ConvNetagainst 77 other previous classifiers. The accuracy of classification is compared with 22 state‐of‐the‐art classifiers. The comparative results were taken from [60]. The classification accuracy and standard deviation results of the tenfold cross‐validation of R‐ConvNet and other methods in previous studies are compared in Table 5. In this table, the methods marked with * use a group of classifiers, each aiming at a specific class, or discriminating among two exact classes in the problem. TABLE 5Comparison of proposed method with classifiers in previous works.CategoryMethodClassification accuracyStd. Dev.EnsembleAdaBoostClassifier0.28890.013BaggingClassifier0.88510.0084GradientBoostingClassifier0.92530.0059RandomForestClassifier0.87270.0140Linear‐modelLogisticRegression0.67660.0199LogisticRegressionCV0.94050.0106PassiveAggressiveClassifier0.70720.0561RidgeClassifierCV0.77060.0239SGDClassifier0.68000.0361Naïve‐BayesBernoulliNB0.77490.0097GaussianNB0.73750.0215MultinomialNB0.20050.0094NeighboursKNeighborsClassifier0.85210.7825NearestCentroid0.51790.4885RadiusNeighborsClassifier0.14260.1499SVMSVC0.13580.0087TreeDecisionTreeClassifier0.81480.0169ExtraTreeClassifier0.55090.0184MulticlassOneVsOneClassifier* 0.95940.0079OneVsRestClassifier* 0.93850.0108OutputCodeClassifier* 0.93110.0100Optimization techniquesEA‐optimized CNN0.96600.0086R‐ConvNet0.99140.0029 Comparison of proposed method with classifiers in previous works. The proposed optimization technique is compared with EA‐optimized CNN [60]. Therefore, the R‐ConvNet result is grouped with EA‐optimized CNN in this table and the results show that meta‐heuristic methods perform better results. Overall, the classification accuracy of R‐ConvNet outperformed other methods in Table 5. The performance analysis is performed using four performance metrics. The performance metrics are precision, f‐measure, accuracy, and specificity which are calculated based on true positive, true negative, false positive, and false negative as follows: True positive (AB):patient diseases are categorized as true as true True positive (AB):patient diseases are categorized as true as true True negative (XY): patient diseases are categorized as true as false True negative (XY): patient diseases are categorized as true as false False positive (DE): patient diseases are categorized as false as false False positive (DE): patient diseases are categorized as false as false False negative (WZ): patient diseases are categorized as false as true False negative (WZ): patient diseases are categorized as false as true The performance metrics, precision, F‐measure, accuracy, and specificity are calculated by Equations (8)–(11) consequently as follows: The performance of the R‐ConvNet is compared with the results of various existing classifiers CMiRNA‐BC‐RF‐SVM, CMiRNA‐BC‐CNN, CMiRNA‐BC‐GCNN, and CMiRNA‐BC‐DCGAN which are all taken from [6]. The results are given in Table 6. This table shows a precision of 0.98% which is higher than other methods compared in this table. The value of F‐measure and accuracy for R‐ConvNet (0.97% and 0.99% for F‐measure and accuracy respectively) is also higher than other methods. The specificity for R‐ConvNet is higher than CMiRNA‐BC‐RF‐SVM, CMiRNA‐BC‐CNN, and CMiRNA‐BC‐GCNN and it provides 0.96% the same as CMiRNA‐BC‐DCGAN. Overall the results in Table 6 demonstrate the high performance of R‐ConvNet compared to other available classifiers when we compare the value of performance metrics. The performance of the R‐ConvNet is compared with the results of various existing classifiers CMiRNA‐BC‐RF‐SVM, CMiRNA‐BC‐CNN, CMiRNA‐BC‐GCNN, and CMiRNA‐BC‐DCGAN which are all taken from [6]. The results are given in Table 6. This table shows a precision of 0.98% which is higher than other methods compared in this table. The value of F‐measure and accuracy for R‐ConvNet (0.97% and 0.99% for F‐measure and accuracy respectively) is also higher than other methods. The specificity for R‐ConvNet is higher than CMiRNA‐BC‐RF‐SVM, CMiRNA‐BC‐CNN, and CMiRNA‐BC‐GCNN and it provides 0.96% the same as CMiRNA‐BC‐DCGAN. Overall the results in Table 6 demonstrate the high performance of R‐ConvNet compared to other available classifiers when we compare the value of performance metrics. TABLE 6Performance analysis of precision, F‐measure, accuracy and specificity with classifiers in previous works.TechniquePrecisionF‐measureAccuracySpecificityCMiRNA‐BC‐RF‐SVM0.610.630.520.59CMiRNA‐BC‐CNN0.590.810.710.77CMiRNA‐BC‐GCNN0.750.600.790.58CMiRNA‐BC‐DCGAN0.960.950.940.96R‐ConvNet0.980.970.990.96 Performance analysis of precision, F‐measure, accuracy and specificity with classifiers in previous works. Table 7 provides the performance analysis of precision, F‐measure, accuracy, and specificity with optimization techniques in previous works. The optimization techniques available in previous works [6] are CMiRNA‐BC‐MMOICA, CMiRNA‐BC‐MSSO, CMiRNA‐BC‐MOO, and CMiRNA‐BC‐MFO. This table shows that the performance of the R‐ConvNet is much better than the first three methods (CMiRNA‐BC‐MMOICA, CMiRNA‐BC‐MSSO, and CMiRNA‐BC‐MOO) and it is comparable with the best‐known method available (CMiRNA‐BC‐MFO) as it provides the same value of 0.99% and 0.96% in accuracy and specificity respectively, the better value (0.97%) in F‐measure and lesser value (0.98%). Since the R‐ConvNet has a better value of F‐measure and F‐measure is more important when we have different numbers of samples for each class, as is the case in this study (Table 1), therefore it can be concluded that the proposed R‐ConvNet has a superior performance compared to other optimization techniques given in Table 7. TABLE 7Performance analysis of precision, F‐measure, accuracy and specificity with optimization techniques in previous works.TechniquePrecisionF‐measureAccuracySpecificityCMiRNA‐BC‐MMOICA0410.490.410.49CMiRNA‐BC‐MSSO0.810.550.810.55CMiRNA‐BC‐MOO0.690.360.690.36CMiRNA‐BC‐MFO0.990.960.990.96R‐ConvNet0.980.970.990.96 Performance analysis of precision, F‐measure, accuracy and specificity with optimization techniques in previous works. CONCLUSION Cooperation between two CNNs was proposed in this paper for age prediction of DNA methylation data. This cooperation was in this manner: at any time only one of the CNNs is active and performs the training process. The activity of training is exchanged between two CNNs periodically. Before each exchange, the knowledge from the currently active CNN is sent to the upcoming active CNN and then this CNN becomes active. Sending knowledge from the source to the target boosts the target and consequently improves the convergence ability periodically. In this method, GA was provided with an optimization process to aid the training phase. During the training process, the initial population is generated. Parents are selected and a combination of single‐point and uniform crossovers is performed and random mutation is operated to form the final children. If the first CNN is active, the first child is fed to the CNN whilst if the second CNN is active, the second child is fed to the CNN. Evaluation of the new model is performed and the best model is updated. This process is continued until a termination criterion is reached. FPPP was the only class of cancer such that the proposed approach is placed in the range of accuracy > 85%. Four of the classes (KIRC, LUAD, PRAD, and UCEC) were in the range of accuracy > 90% and the rest have shown accuracy of 100%. In seven cases (ACC, CESC, CHOL, DLBC, ESCA, HNSC, and MESO) out of 29, the method achieved 100% classification accuracy that no other classifier in previous works could reach. The performance analysis of the results using precision, F‐measure, accuracy, and specificity metrics was performed. The results were compared with the other available classifiers and optimization techniques. In the case of comparison with the classifiers, the method performed better in three cases (precision, F‐measure, and accuracy) and provided the same value for the specificity. In the case of comparison with the other optimization techniques, the R‐ConvNet provides comparable values while being better in F‐measure. Since the F‐measure is more important compared to the other metrics in the case of having different numbers of samples for each class, as is the case in this study, it is concluded that R‐ConvNet performed superior compared to other methods. The high performance of the proposed method is due to using the advantages of deep learning and the relationship between two CNNs. The future of this work firstly concerns a deeper analysis of the R‐ConvNet when applied to other real‐world problems, especially image processing problems in which CNNs show high ability in this subject. In this experiment, the same structure of CNNs in R‐ConvNet was considered. Using different structures of CNNs is another approach for the future work of this study. Furthermore, we aim to apply the proposed method (cooperation between two CNNs) with auto‐control of knowledge transmission between two CNNs which automatically decides the time for the knowledge transmission. AUTHOR CONTRIBUTIONS Najmeh Sadat Jaddi: Formal analysis; investigation; methodology; validation; writing—original draft. Salwani Abdullah: Conceptualization; funding acquisition; writing—review and editing. Say Leng Goh: Writing—review and editing. Mohammad Kamrul Hasan: Writing—review and editing. CONFLICT OF INTEREST STATEMENT The authors declare no conflicts of interest. Jaddi, N.S. , Abdullah, S. , Goh, S.L. , Hasan, M.K. : A related convolutional neural network for cancer diagnosis using microRNA data classification. Healthc. Technol. Lett. 11, 485–495 (2024). 10.1049/htl2.12097 DATA AVAILABILITY STATEMENT The data that support the findings of this study are available from the corresponding author upon reasonable request. REFERENCES Associated Data Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.