A deep hybrid feature extraction model based on MobileNet and vision transformer for acute lymphoblastic leukemia prediction

Abstract

Abstract Early and accurate identification of Acute Lymphoblastic Leukemia (ALL) from peripheral blood smear images can support microscopic screening, but reliable evaluation requires both discriminative feature learning and transparent reporting of the experimental protocol. This study presents MoViTHAFX-Net, a hybrid MobileNet–Vision Transformer (ViT) architecture with Hierarchical Attention Fusion (HAF) and Grad-CAM++ visualization for binary classification of ALL blasts and healthy hematogones (HEM) using the 10,661 labelled cell images in the training component of the C-NMC 2019 dataset. MobileNet extracts local morphological information, whereas ViT models long-range contextual relationships. HAF first refines the two branches and then adaptively combines three learnable morphology-guided feature projections before binary classification. The reported experiment uses an image-level stratified partition containing 7462 training images, 1600 validation images, and 1599 held-out test images; augmentation is restricted to the training partition. On the held-out test partition, the model correctly classified 1583 of 1599 images, corresponding to 99.00% accuracy, with precision, recall, and F1-score of approximately 0.990 and an AUC of 0.99. Ablation results within the same experimental setting indicate an incremental advantage of HAF over direct CNN–ViT concatenation. Grad-CAM++ is used as a qualitative post hoc visualization of image regions contributing to predictions. Because the reported custom split is image-level rather than subject-disjoint and evaluation is limited to a single benchmark dataset, the results should be interpreted as internal benchmark performance rather than evidence of patient-independent or multicentre clinical generalization.