Deep Learning for Medical Image Classification: CNN-Based Approaches Applied to Leukemia and Retinal Disease Detection

Abstract

This thesis investigates the application of transfer learning with pretrained convolutional neural networks for automated medical image classification across two clinically critical domains: acute lymphoblastic leukemia detection from microscopic blood smear images, and binary retinal disease classification from optical coherence tomography scans. Three CNN architectures VGG16, MobileNetV2, and EfficientNetB0 are evaluated under four training strategies: frozen backbone transfer, two-phase fine-tuning, data augmentation, and SVM-based hybrid classification. A total of 24 model configurations are systematically compared under identical controlled conditions on the C-NMC 2019 leukemia dataset and the restructured Kermany OCT dataset. Results show that EfficientNetB0 with two-phase fine-tuning achieves the best performance on both tasks, reaching 98.66% accuracy on OCT and 80.2% on the leukemia dataset, including perfect ALL recall. MobileNetV2 demonstrates competitive results with significantly fewer parameters, making it a strong candidate for resource-constrained deployment. These findings contribute a rigorous and reproducible comparative framework for CNN-based transfer learning in medical imaging