Comparative analysis of deep learning encoders for binary glioblastoma segmentation on post-treatment MRI

Abstract

This study evaluates six encoder backbones within the U-Net architecture for automated whole-tumor segmentation of post-treatment glioblastoma (GBM) using T1-weighted MRI images from the UCSD-PTGBM dataset. The models were compared based on segmentation accuracy, model complexity, and computational efficiency, while three preprocessing strategies were also assessed. Among the evaluated models, EfficientNet-B3 achieved the best overall performance, with a Dice coefficient of 0.763, IoU of 0.618, 13.16 million parameters, and an inference time of 119.60 milliseconds. The study also found that training on raw MRI images produced better results than enhanced images, demonstrating that EfficientNet-B3 with raw-image training provides an efficient and accurate approach for post-treatment glioblastoma delineation.