Deep Learning based Leukemia Detection System using Microscopic Images

Abstract

Blood cancer creates a disruption to healthy white blood cell production in bone marrow; this type of cancer is known as leukemia. Current methods used to diagnose leukemia rely on manually analyzing blood smears, but have a slow turnaround, require a high level of expertise and suffer from inter-observer variability. These limitations create an urgent need for automated, reliable methods for classifying leukemia sub types (ALL, AML, APML, CLL and CML) using a data set called LeukemiaAttri. This study presents innovative approach based on a hybrid deep learning architecture to assist with the automatic classification of five sub types of leukemia using this previously unexamined data set.. Five CNN models and seven Vision Transformer models were trained across two augmentation phases. MobileNetV2 achieved the best CNN accuracy of 95.58% while Swin Base Patch4 Window7 achieved the best ViT accuracy of 95.02%. The two models were fused using Feature-wise Linear Modulation (FiLM), which allows the ViT branch to intelligently modulate CNN features at three scales through learned parameters, enabling richer feature interaction than simple concatenation. A 97% accuracy on the final test set and an AUC-ROC of 0.9993 were achieved with the fused model. The external validation based on 4,915 independent images demonstrated the model's ability to generalise, having an accuracy of 94.85% and AUC-ROC of 0.9949. Grad-CAM++ heat maps confirmed that the model attends to clinically relevant regions of cells, supporting interpretable and trustworthy predictions.The system was deployed as LeukoDetect AI, a Flask-based web application offering real-time classification, confidence scores, and Grad-CAM++ visualizations to support clinical decision making.