Evaluating the impact of region of interest detection methods on medical image classification

Abstract

Medical image classification remains a challenging task due to the subtle and varied nature of disease patterns across imaging modalities. Deep learning models offer promising solutions; however, the integration of region-of-interest (ROI) detection into the training process is still not well understood.This thesis explores the effectiveness of Grad-CAM as an unsupervised ROI detection method within a two-phase framework.In Phase 1, Grad-CAM is used to generate ROI-focused images from chest X-rays and brain MRIs without requirin pixel level annotations. In Phase 2, we train and compare deep classification models using both these ROI- based inputs and the original full images. The architecture consists of a pretrained convolutional backbone (EfficientNetB4 or DenseNet201), a custom classification head, and two fine-tuning strategies: frozen and partially unfrozen (top 25 % trainable layers). Results show that full-image inputs consistently outperform ROI-transformed versions, with DenseNet201 and partial unfreezing achieving the highest accuracy (98.00 % on chest X-rays, 99.00 % on brain MRIs). These findings indicate that while Grad-CAM is valuable for visual interpretation, it may not serve as an effective unsupervised ROI`detector during training, as it`may exclude contextual cues critical for robust learning.