A COMPREHENSIVE COMPARATIVE STUDY OF DEEP LEARNING MODELS FOR BIOMEDICAL IMAGE SEGMENTATION

Abstract

Accurate segmentation of brain tumors in MRI plays a critical role in diagnosis, treatment planning, and disease monitoring. This study proposes a comprehensive comparison of advanced deep learning architectures—3D U-Net, U-Net with ResNet50 Backbone, and Attention U-Net—using the BraTS 2020 dataset for multi-class brain tumor segmentation. The models underwent thorough training with preprocessing steps, custom loss functions, data augmentation, and dynamic learning rate adjustments. Performance was evaluated using Dice Coefficient, Mean IoU, Accuracy, Precision, Sensitivity, and Specificity. Results showed high overall accuracy (0.99) across models, but significant differences in Mean IoU and per-class Dice Coefficients due to class imbalance. The U-Net with ResNet50 Backbone achieved the best Mean IoU (0.64) and balanced per-class performance, highlighting the effectiveness of pre-trained encoders. This work provides insights into model selection and training strategies, paving the way for improved clinical applications.