Rethinking Transfer Learning: Domain-Specific Deep Models for Brain Tumor Detection

Overview

Our project develops a deep learning framework to assist doctors in identifying and diagnosing brain tumors by comparing pre-trained and original models. Using the ResNet50 architecture, our “Original” model was trained from scratch on specific MRI data specific to brain tumors, while the “Pretrained” model leveraged transfer learning from the ImageNet/COCO datasets. Although transfer learning is common in computer vision, its efficacy in highly specialized medical tasks requires further validation. Our experiments showed that our Original model outperformed the Pretrained model in tumor detection, achieving an Intersection over Union (IoU) score of 0.8088 compared to 0.6942. For classification, the Original model achieved 93.75% accuracy, slightly below the Pretrained model’s 98.83% accuracy. These results suggest that for highly specialized medical imaging tasks, training models from scratch to learn domain-specific features can be more effective than relying on pre-learned weights, particularly for object detection.