Blood is an essential element for life and general health of human beings. It performs several crucial functions including transport of nutrients and waste materials, controlling flow of oxygen and overall immune system of body. Human blood is composed of three main types of blood cells, namely, erythrocytes, thrombocytes, and leukocytes. Each cell type performs a specific function in the human body. For example, leukocytes also referred as white blood cells (WBCs) are responsible for human immune and inflammatory response against diseases. Any abnormality in the structure and count of blood cells leads to certain diseases. As an example, leukemia, a blood malignancy, is caused due to an excessive leukocyte production in the bone marrows.

Leukemia is a widespread disease with over 475,000 new cases diagnosed worldwide each year and 312,000 annual deaths (Sung et al., 2021). With 62,770 new cases and 23,670 deaths anticipated, leukemia remains a significant public health concern for the United States in 2024 (Siegel et al., 2024). It is primarily categorized into two types: acute and chronic. Acute leukemia is distinguished by the rapid and unregulated proliferation of immature white blood cells within the bone marrow, which displaces the healthy cells. The fast progression of disease requires prompt response. On the other hand, the chronic leukemia is a slow progressing disease in which gradual accumulation of mature but abnormal WBCs takes place. Although these cells are typically more functional than those found in acute leukemia, they are aberrant and can still affect the normal functionality of blood and bone marrow. The acute and chronic categories of leukemia are further classified into myeloid and lymphoblastic sub-types, based on their afflicted cells. The acute lymphoblastic type of leukemia (ALL) affects the lymphoid cells and has high likelihood of occurring in the children and young adults. It represents ~14% of all new leukemia cases. Approximately 90% of ALL cases occur in individuals younger than 20 years old, with a peak incidence observed in children aged 2–5 (Sung et al., 2021). An estimated 6,550 new cases of ALL are expected in the US in 2024 (Siegel et al., 2024).

A form of acute lymphoblastic leukemia called B-cell acute lymphoblastic leukemia (B-ALL) develops from abnormal B-cell progenitors. Various sub types of B-ALL are further categorized based on distinct genetic, molecular, and immunophenotypic characteristics. Sub types of B-ALL include pre-cursor, mature, common, and pro B cell all.

The classical approach for the diagnosis of leukemia involves visual analysis of microscopic blood images by hematologists. This manual process needs human supervision; therefore, it is a time-consuming process and often prone to classification errors due to several factors (Matek et al., 2019). Thus, an accurate, computer-aided diagnosis of leukemia is highly desirable (Khattak et al., 2022). Among the modern approaches of computer vision, deep CNNs have demonstrated significant potential for a number of classification tasks in the biomedical domain. However, the computer vision-based blood analysis for leukemia diagnosis is difficult due to the small size, irregular structure, and physical similarities across various blood components (Kassani et al., 2019). Moreover, the performance of CNNs depends heavily on their depth and structure. To obtain a high level of accuracy requires a large, accurately labeled dataset for deep neural network training from the scratch. However, due to a number of limitations, such datasets are frequently not easily accessible in the biomedical domain. In such a context, transfer learning stands out as the recommended strategy, entailing the retraining of a deep CNN originally trained on a substantially extensive dataset to suit a specific classification task. A number of pretrained CNNs have achieved high top-1 accuracy on benchmark datasets. GoogleNet (Szegedy et al., 2015), Resnet (He et al., 2016), Darknet (Redmon and Farhadi, 2018), Densenet (Howard et al., 2017), and Inception (Chollet, 2017) are a few to mention. Recent research uses deep CNNs as extractors of features, which are then utilized to train outer classifiers. This leverages the power of transfer learning, allows for task-specific adaptation, and provides an efficient way to build accurate models. However, due to a large number of layers, deep CNNs extract high dimensionality feature representations from the input data. Afterward, feature selection is done to reduce the dimensionality of these extracted features, making them more manageable and potentially more informative. Efforts in current research are directed toward optimizing the computational efficiency and memory demands of the classification pipeline. The primary goal is to attain superior accuracy while operating with a more streamlined feature set (Khan et al., 2020; Ahmad et al., 2023b).

The remainder of the study is structured as follows: Section 2 presents a literature review of some recently published studies in the domain of leukemia identification. Section 3 offers an elaborate exposition of the proposed framework for ALL identification. In Section 4, we present and analyze simulation results, while discussion is concluded in Section 5.

Table 1 presents a summary of some notable contributions in the realm of lekuemia identification using deep learning. They are discussed as follows. In the study mentioned in the reference, Elhassan et al. (2022), an approach is proposed for the detection of acute myeloid leukemia (AML) from WBC images. At first, a CMYK moment-based localization method is proposed to isolate the region of interest (ROI) from WBC images. This is followed by extraction and fusion of several pointwise and spatial features. Classification is performed using multiple classifiers including SVM and XG boost. The study reports the best accuracy of 97.57% on self collected single cell morphological dataset. In the study mentioned in the reference, Dese et al. (2021), a computer-assisted system is proposed for the diagnosis of several leukemia sub-types. The system is based on Gaussian and Weiner filtering for image pre-processing, followed by K-means clustering and marker-controlled Watershed algorithm for segmentation. Several morphological, texture, and statistical features are extracted and classified using multi-class SVM classifier. The best accuracy of 97.69% is reported for overall leukemia detection on self-collected dataset of peripheral blood smear images. In Al-jaboriy et al. (2019), an automatic method for the diagnosis of leukemia is proposed based on leukocyte cell segmentation. The method uses a dataset of 108 microscopic images and performs ANN-based segmentation and extracts various statistical features for classification. The best accuracy of 96% is achieved for binary classification of leukocyte cell blasts. The study mentioned in the reference, Kassani et al. (2019), the authors applied different augmentation techniques to the dataset images. Then, a hybrid CNN model consisting of hidden layers of VGG16, and MobileNet is proposed for feature extraction. The extracted features are classified using a NN architecture. The proposed method achieves a binary classification accuracy of 96.17%. In the study mentioned in the reference Jung et al. (2022), the authors proposed a custom CNN model for WBC classification for leukemia detection. The authors first created a synthetic dataset of WBC images using generative adversarial networks and then performed transfer learning of the proposed CNN for classification. An average accuracy of 97% is achieved by the system.

For extracting and choosing blood features, the authors of the study mentioned in the reference, Alruwaili (2021) presented a stepwise linear discriminant analysis technique. The suggested method performs the identification of specific attributes within blood smear images and their classification based on partial F-values. A Matlab-based method for classifying and identifying WBC cancer was proposed in the study mentioned in the reference, Nithyaa et al. (2021). The approach integrates a range of morphological, clustering, and image pre-processing procedures with the utilization of random forest classification. In the study mentioned in the reference, Pang et al. (2015), an automatic leukocyte categorization approach is proposed. Initially, moment invariants are derived using the Euclidean distance transform within the nucleus region, followed by the extraction of morphological characteristics from the segmented cells.

The published literature on leukemia detection also proposes a number of proprietary deep CNNs and their ensembles. In the study mentioned in the reference, Batool and Byun (2023), a lightweight deep learning-based EfficientNet-B3 model is proposed which employs depth-wise separable convolutions for ALL classification. The method proposed in this study attains a classification accuracy of 96.81% when applied to publicly available datasets for leukemia sub-type classification. In the study mentioned in the reference, Kumar et al. (2020), a simple method for the detection of ALL, and AML is proposed in which KBest algorithm is used for feature extraction, followed by a dense CNN for classification. The proposed approach reports the best accuracy of 97.2%. In the study mentioned in the reference, Jha et al. (2022), a leukemia identification method is proposed which uses K-means clustering from image segmentation. Next, multiple statistical features are extracted to train an ensemble of multiple classifiers. The proposed system reports a best accuracy of 96.3%. In most of the existing studies that utilize deep transfer learning, the feature selection is performed using a filter or wrapper-based approach. Filter-based methods assess the relevance of individual features by examining their statistical properties, such as correlation with the target variable or variance within the feature. These methods have a limitation in that they do not consider the relevance between the selected features and the actual model's performance. This can lead to situations where selected features might not be the most predictive for the planned model. Conversely, wrapper-based methods entail employing a machine learning model in the capacity of a “wrapper” to assess the effectiveness of various feature subsets. These methods select features by repeatedly training and evaluating the model on different subsets of features. These methods are particularly useful in obtaining the best set of features for a specific classifier model. Recently, population-based algorithms for feature selection have received considerable research attention. A significant challenge lies in fine tuning the algorithm to achieve better exploration of feature search space and obtain the most discriminant and powerful set of features. Standard population-based algorithms used in several studies on disease classification often suffer from poor convergence and local optima problems (Gupta et al., 2020; Shahzad et al., 2022).

The prescribed workflow for the detection and sub-type categorization of acute lymphoblastic leukemia has been executed using MATLAB R2021a, running on an Intel Core i7 CPU equipped with 16GB of RAM, all hosted within a 64-bit Windows 10 operating environment.

In this study, we examined the effectiveness of our proposed approach for the binary and multi-class identification of ALL. Modern deep CNNs often come with large model sizes, demanding significant memory and computational resources. Employing an ensemble of networks, including a tailored CNN alongside publicly available deep CNN models, offers a practical compromise between classification performance and pipeline complexity. Furthermore, leveraging pretrained CNN models for feature extraction and employing external classifiers is a potent and pragmatic strategy that amalgamates the advantages of transfer learning, feature abstraction, and minimized training effort to enhance the outcomes of diverse computer vision tasks. One drawback of this approach is that the feature sets extracted from deep CNNs often exhibit substantial size and encompass a considerable amount of duplicate features. Selection of most promising set of features is a combinatorial optimization problem with computational complexity of exhaustive search growing exponential with the size of feature vector. Population-based feature selection methods have shown a significant research interest in recent years. A number of bio-inspired and nature-inspired meta-heuristics have been proposed. One challenge is the exploration and exploitation capabilities of the algorithm problem of local optima. To address this issue, we have proposed a memetic feature selection approach that combines elements of population-based algorithms with local search methods. In particular, we have proposed a nature-inspired metaheuristic name binary whale optimization algorithm in which optimization of an iteration best solution is performed using a differential evolution method. These optimizations at CNN architecture and feature selection level yield an improved pipeline which shows promising results for leukemia detection and sub-type classification. The validity of proposed approach is manifested with better performance results as compared with several recently published studies.

Leukemia, a hematologic malignancy, afflicts both pediatric and geriatric populations. Acute lymphoblastic leukemia is an aggressive form of leukemia that has a high mortality rate. Modern computer vision approaches and deep CNNs have been demonstrated as potential solutions for computer aided diagnosis of several medical conditions. However, precise classification of malignancies at microscopic level is a challenging task due to morphological similarities between different blood entities. This study presents an improved pipeline for enhancing leukemia detection from blood smear images. At first, We propose an intricately designed 88-layer deep CNN architecture inspired by AlexNet and SqueezeNet. We used this network as a feature extractor alongside GoogleNet, aiming to balance classification accuracy and computational efficiency. The work then models the feature selection problem as a combinatorial optimization problem and proposes a novel memetic approach based on the Hybrid binary whale optimization algorithm to meticulously select the most dominant set of features. Our proposed methodology undergoes rigorous validation using publicly available datasets containing peripheral blood smear images across diverse leukemia classes. The proposed feature selection approach effectively selects the most dominant and discriminant set of features. The proposed system achieves an overall accuracy rate of 99.15% with an 80% reduction in feature size, performing comparably or better than several existing studies on leukemia identification. The propose method can be extended to the diagnosis of other blood-related diseases. It can complement advanced diagnostic methods such as RNA sequencing and molecular testing by providing additional supporting evidence. Additionally, it offers smooth integration with practical image analysis systems such as image flow cytometry, expanding their functionalities in real-world settings.

Publicly available datasets were analyzed in this study. This data can be found here: Scotti et al. (2005) and Ghaderzadeh et al. (2022).

MA: Conceptualization, Data curation, Investigation, Methodology, Software, Validation, Writing – original draft, Writing – review & editing. RA: Methodology, Software, Validation, Writing – original draft. NK: Supervision, Writing – review & editing. AA: Writing – review & editing. NA: Writing – review & editing. AM: Resources, Validation, Writing – review & editing.