Enhancing Brain MRI Classification Through Layer-Wise Transfer Learning and Deep Fine-Tuning

Abstract

Brain tumors must be correctly and quickly diagnosed in order for treatment to work and for patients to have a better chance of life. MRI is a very important way to see the shape of brain tumors. However, diagnosing them by hand takes a long time and can be inaccurate depending on who is doing it. We looked at three different versions of the DenseNet201 model to see how well they work at automatically putting brain tumors into four groups: glioma, meningioma, pituitary tumor, and no tumor. The first variant, B-DenseNet201-CA, is trained entirely from scratch to establish a baseline. The second variant, PT-DenseNet201-FB, employs transfer learning with a frozen DenseNet201 base, leveraging pretrained ImageNet features while training only the classification head. The third and most advanced variant, FT-DenseNet201-TL, fine-tunes the last 50 layers of the pretrained DenseNet201 to enhance domain-specific adaptability. A standard MRI dataset was used to train and test all the models. Confusion matrices, accuracy, precision, recall, and F1-scores were used to look at how well they did. The baseline model was only 74.11% accurate, which shows how hard it is to learn from scratch in medical imaging tasks with little data. The frozen transfer learning model significantly improved accuracy to 90.81%, demonstrating the utility of pretrained features. The fine-tuned transfer learning variant further enhanced performance, reaching an accuracy of 94.18% and exhibiting superior class-wise discrimination, particularly for overlapping tumor types. This study shows that partial fine-tuning of a pretrained DenseNet201 is a strong, computationally efficient, and very accurate way to diagnose brain tumors. It also shows how important transfer learning is in medical picture classification. The results could be used in real-life diagnostic tools to help doctors make decisions more quickly and consistently.