Brain Tumor Segmentation: Swin-UNETR vs. nnUNet vs. U-Net

BraTS 2023 (100 Samples)
AMS 691: Medical Image Analysis, Stony Brook University | Spring 2025

---

Abstract

Benchmarked Swin-UNETR (transformer), nnUNet, and U-Net on 100 BraTS 2023 MRI volumes using MONAI.
Trained on T1, T1ce, T2, FLAIR with DiceCE loss.
Evaluated accuracy, boundary precision, inference time, memory.

nnUNet: Best accuracy (Dice 0.152, HD95 40.79)
Swin-UNETR: Competitive (Dice 0.1399) but 25× slower
U-Net: Fastest (0.05s) but weak (Dice 0.1021)

Open-source pipeline for low-data, low-resource clinical AI.

---

Results (5-Fold Cross-Validation)

Model	Dice ↑	HD95 ↓	ASD ↓	Inf. Time ↓	Memory
nnUNet	0.152	40.79	17.12	0.87s	1.57MB
Swin-UNETR	0.1399	53.83	20.85	1.24s	1.57MB
U-Net	0.1021	100.70	48.31	0.05s	1.57MB

---

Run the Code

pip install -r requirements.txt
jupyter notebook Tumor-Segmentation-BraTS.ipynb