Hybrid YOLO–U-Net Framework for Uncertainty-Aware Brain Tumour Segmentation and Web-Based Visualisation of MRI Scans.
Year: 2026
Type: article
Abstract: Brain tumour segmentation is a challenging task in magnetic resonance imaging (MRI) that demands precise boundary definition and correct understanding of automated delineation. This study introduces a novel YOLO–U-Net system that combines tumour localisation, segmentation, uncertainty quantification and web-based visualisation. The method details an applied experimental design based on de-identified MRI scans, partitioning of patient-level data, intensity normalisation, resizing, and augmentation during training. YOLO detects the candidate tumour regions and U-Net generates the segmentation masks. Monte Carlo dropout makes multiple predictions to estimate uncertainty in detected regions. An original image, segmentation overlay and uncertainty maps are presented using a Streamlit interface. Evaluation is based on Dice similarity, intersection over union, precision, recall, calibration, error detection, and processing time. System comparisons resulted in Dice coefficients of 0.920 for the deterministic hybrid, and 0.917 for the uncertainty-aware hybrid. The stochastic pass performance was enhanced from 10 to 30, which resulted in an AUROC increase from 0.821 to 0.843 for the system error-detection problem, and inference time went from 2.40 to 4.80 seconds. These are synthetic values, testing for potential evaluation trade-offs and not testing for experimental effectiveness. The framework is a structured way to integrate segmentation and reliability information in an easy-to-use interface. Its accuracy, uncertainty quality, computational practicality and likely usefulness to support expert MRI review workflows will need to be determined through independent patient-level experiments, external validation, and usability assessment.
Institution: Caleb University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7236819308
