Integrating convolutional neural networks for microscopic image analysis in acute lymphoblastic leukemia classification: A deep learning approach for enhanced diagnostic precision

Abstract

Accurate and timely classification of acute lymphoblastic leukemia (ALL) based on microscopic blood smear images is a major priority area in computer-aided diagnosis, but the high intra-class similarity and the uneven or variable stains often make the prior models generally limited in their applicability. A hybrid CNN-ViT model is proposed in the study to combine spatial local representations of Convolutional Neural Networks (CNNs) with global contextual relationships of Vision Transformers (ViTs). To alleviate the scarcity of data and the class imbalance, a lot of data augmentation and transfer learning were done, and model explanations were improved by gradient-weighted Grad-CAM to interpret the discriminative cellular regions. A publicly available leukocyte dataset evaluation shows that the proposed Hybrid CNN-ViT achieved a state-of-the-art validation of 99.60%, beating known benchmark models including DenseNet-121, ResNet-50, and EfficientNet-B0. The comparative analysis proves the strength of the model, and visual interpretations demonstrate that the trained maps of attention reflect the biological salient features of leukocytes. These findings point to the clinical promise of hybrid representation learning in the diagnosis of hematological images, notably, to the support of a diagnostic workflow in ALL. Nevertheless, the framework introduces computational overhead and data set specificities, making it necessary to be further validated within a multi-institutional and real-world clinical setting.