Multi-Modal MRI Brain Tumor Segmentation: Benchmarking Volumetric Deep Learning Architectures
编号:106
访问权限:仅限参会人
更新:2026-11-10 16:46:49 浏览:12次
Online
摘要
Segmentation tasks of brain tumour from MRI scans are mostly difficult due to irregular tumor boundaries, intensity variations, and structural heterogeneity across patients. While deep learning has advanced segmentation accuracy, reliably distinguishing tumor tissue from healthy brain regions remains an open challenge. This study presents a comparative evaluation of five 3D segmentation architectures — Federated VNet3D, Federated ResVNet3D, ResUNet3D, SmallUNet3D, and AttentionUNet3D — trained on the BraTS 2020 dataset. Models are assessed using Dice Score, IoU, Sensitivity, Specificity, Precision, F1-Score, Volume Similarity, PSNR, SSIM, and Detection Rate. Among the trained models, Federated VNet3D has shown the strongest results, recording a value of 0.7892 and 0.6740 for Dice Score and IoU respectively, reflecting better structural consistency. An interactive Gradio-based visualization interface was also built to support segmentation overlay and heatmap based tumor localization, improving clinical interpretability.
关键词
Brain Tumor Segmentation,VNet3D,ResVNet3D,Multi-Modal MRI,Federated Learning,Gradio Visualization
稿件作者
Sathish J
Anna University
Varalakshmi P
Anna University
发表评论