Improving acute lymphoblastic leukemia diagnosis through CBAM-enhanced VGG19 deep learning

Acute Lymphoblastic Leukemia (ALL) is an aggressive blood cancer that requires rapid and accurate diagnosis. Manual review of blood and bone marrow smears is time-consuming and observer-dependent, underscoring the need for automated methods. This study presents a deep learning framework with attention mechanisms for automated detection and subtyping of ALL from microscopic bone marrow images, including healthy samples. The model combines a Convolutional Block Attention Module (CBAM) with a VGG19 backbone, forming a hybrid CBAM-VGG19 network that hierarchically enhances key morphological features across spatial and channel dimensions. Incorporating CBAM into VGG19 improves feature extraction, accelerates learning, and boosts classification accuracy, particularly for visually similar leukemia subtypes. The model's performance was validated using k-fold cross-validation, achieving 98.73% classification accuracy, surpassing DenseNet121, InceptionV3, MobileNetV2, and the original VGG19. Further analyses examined the effects of image resolution, hyperparameter optimization, Bayesian tuning, and CBAM layer placement, all of which improved convergence and robustness while reducing overfitting. Although the results are promising, the lack of external validation and the small dataset size limit clinical applicability. Therefore, this framework is a research prototype intended to serve as a basis for future large-scale, multi-center studies.