Computer vision
TumorLytix: unsupervised brain tumor detection and segmentation
CycleGAN plus diffusion on 65,072 MRI images
- Timeline
- Oct 2024 to Dec 2024
- Affiliation
- WPI Deep Learning
- Team
- Ankit Gole, Devesh Bhangale, Farhan Seliya, Sarthak Mehta
- MRI images
- 65,072
- Segmentation accuracy
- >90%
- Precision over baseline
- +15%
- Inference throughput
- 1.6x
What it is
An unsupervised medical-imaging pipeline that pairs CycleGAN image synthesis with diffusion-based refinement to detect and segment brain tumors across 65,072 BraTS2020 MRI images with more than 90% segmentation accuracy and no manual labels.
Problem
Supervised tumor segmentation needs large labeled MRI datasets that are expensive to produce, while tumors vary widely in size, shape and position.
Approach
Train CycleGAN on unpaired scans to synthesize paired healthy and diseased images and flag anomalies where reconstruction fails; refine boundaries with a conditionally guided diffusion model; automate labeling with OpenCV morphological operations and entropy-based filtering.
Highlights
Processed 23,180 healthy and 41,892 diseased slices across T1, T2, T1CE and FLAIR modalities from 369 patients.
Automated labeling improved baseline precision 15% and stabilized performance across modalities.
Packaged TorchScript batched mixed-precision inference for 1.6x images per second, profiled with Nsight Compute.
Synthetic pairing reached 88.2% accuracy on tumor image creation and 73.7% on healthy reconstruction over 75 epochs; versus a plain CycleGAN, 25% fewer artifacts, 30% higher precision and 50% faster convergence.