A Hybrid Framework for Acute Lymphoblastic Leukemia Identification Utilizing VIT-CNN Fusion and Immunologically Inspired Deep Feature Selection

Acute Lymphoblastic Leukemia (ALL) is a serious blood cancer characterized by the abnormal growth of progenitor white blood cells, which interferes with normal blood cell production. Early and precise detection is essential for effective treatment and better patient outcomes. Convolutional neural networks (CNNs) have shown significant promise in automating diagnostic processes within digital pathology. However, classifying ALL subtypes remains challenging due to subtle morphological differences in lymphoblast features compounded by small and imbalanced datasets. This paper presents a hybrid framework for improved ALL classification from blood smear images, leveraging the strengths of deep learning and immunological optimization. The proposed approach integrates Vision Transformer (ViT-B/16) and a customized CNN as feature extractors to capture both global and local representations of leukemia-related patterns. The extracted features are fused and optimized using an immunological clonal selection algorithm, inspired by the adaptive immune system’s mutation and selection process. The reduced set of best selected features is subsequently utilized to train various baseline classifiers with diverse kernel configurations. To rigorously validate the generalization ability and robustness of the proposed approach, this study leverages well-established datasets that include blood smear images from a broad spectrum of ALL classes, ensuring a comprehensive assessment across various conditions. The proposed method achieves an average accuracy of 98% for binary ALL classification, with a 60%reduction in the feature vector, alongside 97.9% precision and 98% sensitivity. For B-ALL subtype classification, it reaches a maximum accuracy of 98.7%, with 98.8% precision and 97% sensitivity. Overall, the proposed approach surpasses several existing methods in terms of key performance metrics.