Advanced Transfer Learning Framework for the Automated Detection and Classification of Acute Lymphoblastic Leukemia

Cancer was historically one of the most malevolent diseases, and despite several treatment options, it remains one of the deadliest to human life. A mutation in the DNA of an individual cell that undermines its function triggers leukemia, resulting in the excessive generation of immature white blood cells that invade the space necessary for the maturation of healthy blood cells. Conversely, leukemia or blood cancer is incurable if diagnosed and treated at an advanced stage. The diagnosis of blood cancer is a laborious, comprehensive, and vital process that is nearing failure. In this analysis, we have introduced an innovative method for diagnosing blood cancer using white blood cell pictures with the application of Deep Learning (DL). This study utilized the publicly available ”Acute Lymphoblastic Leukemia (ALL) image dataset,” comprising 3,256 stained images of white blood cells and 3,256 expertly labeled segmented images. To augment and balance the training data, the Synthetic Minority Oversampling Technique (SMOTE) was employed. We proposed a compact dataset, as our objective was to enable the model to train effectively on limited data, achieving significant accuracy from the short dataset. We utilized many Convolutional Neural Network models to extract features and developed a bespoke head for leukemia classification. We employed DenseNet121, InceptionV2, MobileNetV2, ResNet152V2, VGG16, ViT, and our proprietary advanced Custom Model. All models have exhibited promising outcomes; nevertheless, despite ResNet152V2 attaining the highest accuracy of 99.69%, we have introduced an ensemble method. The application of the soft voting technique produces a final output with 99.99% accuracy in the diagnosis and classification of leukemia. GRAD-CAM (Gradient-Weighted Class Activation Mapping) was utilized to ensure the model’s convergence. Thus, the methodology provides a reliable ensemble model for the swift detection of leukemia in blood smear images.