Brain Tumor Segmentation Using Residual 3D U-Net on the BraTS 2020 Dataset
"BraTS 2020 based deep learning model for accurate brain tumor sub‑region segmentation with strong false positive suppression."
Problem Statement
Brain tumor MRI segmentation is difficult due to severe class imbalance, complex tumor shapes and high false positives. Manual segmentation is slow and subjective. We design a Residual 3D U‑Net model that improves stability, accuracy and clinical reliability.
Literature Review / Market Research
Analyzed U-Net, Attention U-Net, Transformer models, BraTumIA pipeline, and commercial tools like Siemens syngo.via and GE Edison. Most advanced models are complex and need large proprietary datasets.
Research Gap / Innovation
Existing academic models show unstable training and high false positives. Our work improves reliability using residual learning, Focal Tversky loss, and post-processing without increasing complexity.
System Methodology
Dataset / Input
BraTS 2020 Dataset (100 patients subset). Multi-modal MRI: T1, T1ce, T2, FLAIR. Patch size: 96×96×96 with normalization.
Model / Architecture
Baseline: 3D U-Net. Proposed: Residual 3D U-Net (ResUNet v2). Loss: Focal Tversky Loss. Training: Adam optimizer, cross-validation, post-processing using connected component filtering.
Live Execution
VIEW CODE / DEMOResults & Analysis
ResUNet v2 improves Dice scores across all tumor regions. Enhancing tumor Dice reached ~0.83–0.85 with high precision and strong false positive suppression compared to baseline 3D U-Net.
Academic Credits
Project Guide
Dr. Amit Kumar Gupta
Team Member 1
Manasvi Walia
23FE10CSE00091