This study used a set of data from the Kaggle repository [41]. The photos in this collection were created in the bone marrow laboratory of Taleqani Hospital (Tehran, Iran) and include 3262 genuine peripheral blood sample images. These photos were obtained from 89 patients, 25 of whom were recognized as fit, whereas the other 64 were confirmed with ALL. The set of data is classified into two main categories: malignant and benign. The malignant group is further split into three subgroups: early, pro-B, and pre-B. The photographs were taken with a Zeiss camera fitted to a microscope at 100 × magnification and saved in the form of a JPG. A professional used the technique of flow cytometry to precisely classify these photos. Figure 2 depicts one of the sample photos utilized in this investigation. The dataset used in this work was divided into two parts: train (80%) and test (20%) based on patient-wise split. This strategy ensures that images belonging to the same patient do not appear in both the training and testing sets. This approach prevents information leakage and ensures a more realistic evaluation of model performance. Table 2 provides an overview of the entire dataset.

Figure 1.Overall architecture of the proposed system.

Table 2. Working dataset distribution.

The data pretreatment pipeline consists of various processes that guarantee the input photos are appropriately structured for further examination. First, the photos are loaded into OpenCV, and any illegible ones are discarded. Each picture is then transformed to grayscale, which reduces complexity yet retains fundamental integrity. Otsu’s thresholding approach is used to separate both foreground and background objects. This approach provides a binary picture in which the item appears in white and the background is black. The foreground pixels are subsequently recognized so that only the area of interest remains. If no foreground pixels are identified, the picture is discarded. The captured item is cropped to eliminate any superfluous background areas. After cropping, the picture is enlarged to 224 × 224 pixels to ensure uniformity in the input dimension. This preliminary processing method improves the overall quality of the data being input by removing noise, standardizing image parameters, and focusing on specific areas. We handle the grayscale images for training the CNN architecture through image resizing and normalization steps. Firstly, all images are resized to 224 × 224 pixels to match the input requirements of the CNN architectures. Then, the pixel values of each image are normalized to the range [0, 1] prior to feature extraction.

In this research, four DL-based feature extractors (ResNet50, VGG19, MobileNet, and VGG16) are applied to retrieve the DL features from the blood microscopic images. In the feature-extraction setup, features are extracted from the dataset by following different steps so the feature vectors are reproducible. Firstly, all CNN backbones were initialized with ImageNet pre-trained weights. Then, the convolutional base of each network was used as a feature extractor. After that, the final classification layers were removed, and features were extracted from the Global Average Pooling (GAP) layer. During feature extraction, the convolutional layers were kept frozen to preserve the learned representations and reproduce the feature vector. Each DL model extracts a different number of features, such as the VGG model extracts 512 features, ResNet50 extracts 2048 features, and MobileNet extracts 1024 features. Algorithm 1 shows the feature extraction procedure of each model. The description of each DL model is given in the next subsection.

Algorithm 1. Feature extraction procedure using DL models

Input: 2D microscopic data

Output: DL Feature Map

Initialization:

1. p = 2P - 1, for P = 1, 2, 3... ... .... ... ... p

2. D←Input data

3.X_p ←Use the median filter on the input data D with kernel size p× p

4.M_f ←Extracted Feature Map

Start:

1. For P = 1 to n:

2. CalculateX_p

3. Apply (D,X_p) to findM_f|M_f{R₀,R₁, .....R₁₄}

4.M_f ← M_f

5. end for

6. displayM_f

End

1) VGGNet

VGGNet [42] is a CNN classifier developed by the Visual Geometry Group at the University of Oxford in 2014. This architecture employs a series of 3 × 3 convolutional layers stacked deeper compared to earlier architectures. This architecture uses max-pooling layers to sequentially minimize spatial dimensions while enhancing characteristic richness. It utilizes fully connected layers at the end for classification. In this work, we utilized two VGGNet architectures named VGG16 and VGG19 to extract features, which have 16 and 19 weight layers, respectively. These two networks extract 512 features separately from the input data.

Figure 2.Example of some microscopic images.

2) ResNet50

The ResNet50 [43] model comprises 50 distinct layers with 2M variables. It consists of several components: 64 kernels, convolution level, dense layer, and max-pooling level. The residual block of a ResNet50 model allows for the deterioration issue and removes the vanishing issue. Furthermore, the skip connection block works as a super pathway. In this work, ResNet50 takes ALL images as input and extracts 2048 DL features using the last layer.

3) MobileNet

MobileNet [44] is a lightweight network that is commonly applied in embedded strategies for diagnostic-based systems. The DC (depth-wise convolutions) allows this network to minimize the training time. The operation of the MobileNet network is first the DC, followed by the PC (point-wise convolution). The convolution process of the MobileNet is expressed by the following Equation (1).

here, T indicates the input tensor, K indicates the kernel, T_j denotes the tensor’s j-th component, and * indicates the CO (convolution operation). However, after performing the component-wise product and moving K over T in the convolutional layer, the final result of the CO is calculated by combining T and K. However, this experiment applies the MobileNet DCNN model as the first feature extractor that retrieves 1024 high-impact features.

The operating technique of the offered feature selection method is described in this section. In this experiment, an updated feature selection algorithm named Analysis of Variance Feature Selection (ANOVA) is utilized to update the execution time and predicted results. The description of this algorithm is given below.

Analysis of Variance Feature Selection (ANOVA)

ANOVA [45] is an updated statistical feature selector that ranks characteristics by computing the variance ratios across and within categories. The ratio displays how closely the δ th characteristic is related to the collective attributes. The ratio R for two working datasets is calculated by Equation (2).

where s B 2 ( δ ) and s W 2 ( δ ) are the sample variances between classes and within classes. The formulas for these two sample variances are given in Equation (3) and Equation (4).

the degrees of these two sample variances are defined as d f B = k − 1 and d f W = N − K , where K indicates the number of classes and N is the total number of samples. The frequency of the δ th characteristic in the j th instance in the i th class is indicated by f i j ( δ ) . However, the sum of all examples in the j th class is indicated by n i . The working principle of the ANOVA feature selector is given in Algorithm 2.

Algorithm 2.Feature selection mechanism using ANOVA

Input: Extracted feature map

Output: Optimal feature map

initialization:

1.Y = F-1, for F = No. of retrieved features of each DL algorithm.

2. L←Data labels

3.Y_n ←Training samples

4.L_n ←Training labels

5.M_f ←Respective optimal feature map

6.G_f ←Number of generated optimal feature maps

start:

1. feature = ANOVA Feature Selector(Y_n, L_n)

2. if (Extracted features > 0.5) :

3.G_f= features

4.M_f= sum (total no. ofG_f)

5. end if

6. displayF_v

end

This part represents the final classification tasks utilizing several ML classifiers. Four ML classifiers (SVM, RF, KNN, and NB) are used to classify blood cancer types. Among them, the SVM classifier provided the best results. The working mechanism of this classifier is given below.

Support Vector Machine (SVM)

SVM [46] is an ML paradigm capable of classifying diseases by analyzing input data. It works by constructing an optimal hyperplane that separates data points into different classes within an n-order (D) space, which makes it easier to classify new instances. In this work, the classification process leverages the kernel trick (KT), a method that enables SVM to handle non-linear data. Specifically, for a two-dimensional dataset that is not linearly separable, the kernel trick maps the instances into a higher-order space where a linear partition becomes feasible. The corresponding mathematical expression is provided in Equation (5).

This part reflects the simulated results of the suggested work without feature selectors. We summarized the classification results of each DL model with four ML classifiers from Table 3-6. Table 3 demonstrates simulated results of the VGG16 feature extractor with different ML classifiers. From Table 3, the SVM classifier provides the best results with an accuracy of 98.00%, and the NB classifier provides the lowest performance with an accuracy of 80.58%.

Table 3. The overall performance of different classifiers with VGG16.

Table 4 demonstrates the classification outcomes of the VGG19 feature extractor with different ML classifiers. From Table 4, the SVM classifier provides the best results with an accuracy of 97.69%, and the NB classifier provides the lowest performance with an accuracy of 87.06%.

Table 4.Simulated results of four classifiers with VGG19.

Table 5 demonstrates the classification results of the MobileNet feature extractor with four ML classifiers. From Table 5, the SVM classifier provides the best results with an accuracy of 94.61%, and the NB classifier provides the lowest performance with an accuracy of 62.25%.

Table 5. Simulated results of four classifiers with MobileNet.

Table 6 shows the classification results of the ResNet50 model with different ML classifiers. From Table 6, the SVM classifier provides the best results with an accuracy of 99.54%, and the NB classifier provides the lowest performance with an accuracy of 83.82%.

Table 6. The overall performance of different classifiers with ResNet50.

From the above comparison table, we see that the ResNet50 feature extractor with the SVM classifier obtained the best results among the feature extractors. ResNet50 with SVM classifier (ResNet50 + SVM) provided an accuracy of 99.54%, an F1-score of 99.52%, a recall of 99.56%, and a precision of 99.49%. Table 7 provides a summary of the ResNet50 feature extractor after extracting the features from the input data. The performance metrics of the ResNet50 + SVM network are tabulated in Figure 3. Figure 4 shows the evaluation chart of accuracy from different feature extractors with the SVM classifier.

Figure 3.Performance matrix of the ResNet50 + SVM model: (a) confusion matrix, (b) AUC curve, (c) ROC-AUC curve.

The following parameters are used in each ML classifier to classify different subtypes of blood cancer. In the KNN ML classifier, the best nearest neighbor is K = 14. In SVM, we apply grid search as a cross-validation (CV) technique to find the optimal solution. The best parameters of the SVM classifier are C = 1, gamma = 0.1, and kernel = polynomial. In the RF classifier, the best parameters are max depth = 7, and number of estimators = 200. In this research, we use the Gaussian Naïve Bayes variant to find the optimal solution.

Table 7.Summary of the input layer with output shape and parameters of the ResNet50 model.

Total params:23,587,712 (89.98 MB); Trainable params: 23,534,592 (89.78 MB); Non-trainable params: 53,120 (207.50 KB).

Figure 4.Comparison of the accuracy of different feature extractors with the SVM classifier.

This part presents the experimental outcomes using a feature selection approach named Analysis of Variance Feature Selection (ANOVA). This feature selection algorithm is applied to the best model (ResNet50 + SVM) to improve the classification results of this model. The ANOVA feature selector selects 1678 best features from the extracted feature set. Table 8 shows the corresponding results of the ResNet50 + SVM model after applying the ANOVA feature selector. Table 8 reflects that the feature selector method improved the previous accuracy from 99.54% to 99.69%. The evaluation matrices of the ResNet50 + ANOVA + SVM network are tabulated in Figure 5. Table 8 reflects the corresponding comparison between the two classification methods. The ANOVA algorithm selects 1678 best features from the extracted 2048 features. In the feature selection technique, the proposed framework makes classification results based only on the best features rather than all features. That is why the ResNet50 + ANOVA + SVM model provides better results than the ResNet50 + SVM model.

Table 8.Performance comparison of pneumonia prediction using the ANOVA feature selector for the ResNet50 model.

Figure 5.The performance measurement ANOVA feature selection with ResNet50+SVM (a) Confusion matrix (b) AUC-ROC curve (c) ROC curve for all classifiers.

Validation Protocol

To ensure the robustness and reproducibility of the results, we employed a 10-fold stratified cross-validation protocol on the training set. Unlike standard shuffling, stratification ensures that the class distribution of ALL subtypes in each fold mirrors the original dataset, preventing bias in smaller sub-classes. For each of the K = 10 iterations, nine folds were used for feature extraction and classifier training, while the remaining fold served as the validation set. Table 9 shows the reporting stability of this work with Mean ± Variability score. Table 10 shows the comparative analysis for the SOTA approach on the ALL dataset.

Table 9.The stability reporting of this work with Mean ± Variability score.

Table 10.Comparative results for the SOTA approach on the ALL dataset. The highest outcomes are shown in bold. Here, A stands for accuracy.