Impact of You Only Look Once (YOLO)-Derived Regions of Interest on U-Net Segmentation of Post-Treatment Glioma in Magnetic Resonance Imaging: A Methodological Study on the BraTS (Brain Tumor Segmentation) 2024 Dataset

Abstract

Hybrid deep learning pipelines combining You Only Look Once (YOLO) for lesion detection and U-Net for segmentation are increasingly applied to brain tumor MRI, yet the specific contribution of region of interest (ROI) cropping to segmentation performance remains poorly quantified. This study evaluates four ROI-based U-Net configurations for post-treatment glioma segmentation on the Brain Tumor Segmentation (BraTS) 2024 - GLI dataset, systematically varying the YOLO detector (YOLOv8m vs.YOLO11m) and the U-Net encoder backbone (ResNet-34 vs. ResNet-50). Normalized T1 volumes and expert tumor masks were converted into axial 2D slices. Two YOLO detectors were trained for single-class tumor localization; their bounding-box predictions were used to crop 256 × 256 ROIs with adaptive padding. U-Net models with ResNet-34 and ResNet-50 encoders were then trained on each ROI set using combined Dice and binary cross-entropy loss. Performance was measured by Dice score and intersection over union (IoU) on held-out validation subjects. The best-performing configuration, YOLO11m-U-Net with a ResNet-34 encoder, achieved a validation Dice of 0.8824 and an IoU near 0.80. YOLOv8m-based ROIs with ResNet-34 produced comparable results (Dice 0.8797). ResNet-50 encoders yielded lower scores across both detectors (Dice 0.8656 and 0.8601). Failure cases were predominantly linked to missed or poorly localized detections. These findings indicate that moderate-depth encoders paired with high-quality ROI generation offer the most favorable accuracy-efficiency tradeoff for post-treatment glioma segmentation in hybrid computer-aided diagnosis systems.