A Unified Deep Learning Framework for MRI Brain Analysis: Integrating Segmentation, Classification using Transfer Learning and Generative Models

• Published in: Indian journal of science and technology Vol. 18; no. 32; pp. 2646 - 2655
• Main Authors: Amsavalli, S, Saminathan, K, Devi, M Chithra
• Format: Journal Article
• Language: English
• Published: 23.08.2025
• ISSN: 0974-6846; 0974-5645
• DOI: 10.17485/IJST/v18i32.1265

Objectives: To design a unified deep learning framework that combines brain MRI segmentation and classification through transfer learning and generative models to increase rate of diagnosis, compensating for data imbalance, and improving spatial localization of brain lesions. Methods: The study employs preprocessing of input data and uses DCGAN for augmentation of the brain MRI datasets to cope with class imbalance. Then fine-tuned with a pre-trained CNN (ResNet50, for example) for tumor classification and implemented a U-Net++ with ResNet encoder for segmentation. These tasks were integrated under a multi-task learning framework with a shared encoder and two decoders. Finally, the performance is assessed through cross-validation by using metrics such as accuracy, F1-score, Dice coefficient, and Hausdorff distance. Findings: The unified deep learning framework encompassed in this thesis achieved best performance with classification accuracy of 93.2% and Dice score of 0.91. The use of a shared encoder with two decoders allowed the model to undertake multi-task learning which enhanced model robustness. Furthermore, the proposed method addressed the class imbalance appropriately, particularly with use of DCGAN healthcare dataset images. The proposed method demonstrated more spatial accuracy and interpretability than some of the models already proposed in literature (e.g., M-Net, TransBTS, GAN-AttnNet). Novelty: The proposed study is a general unified multi-task learning framework that can perform MRI brain tumor classification and segmentation through shared representations at the unified image level. Additionally, transfer learning and data augmentation derived from a DCGAN model has been incorporated to improve performance of the classification and segmentation tasks when training with limited and imbalanced medical datasets. Keywords: Brain MRI, Deep Learning, Tumor Classification, Image Segmentation, Transfer Learning, DCGAN, Multi-task Learning, Medical Image Analysis, Attention Mechanism, Generative Models