Self-Supervised 3D Cancer Cell Segmentation
Segmentation 3% better than expert-annotated masks and 20% higher downstream diagnostic accuracy. Published on arXiv and under review at Pattern Recognition (Elsevier).
Challenge
A medtech company needed to segment cancer cells in very large 3D light-sheet microscopy scans, where expert annotations are slow and expensive.
Approach
Designed Barlow-Swin: a Swin-Transformer encoder with a U-Net decoder, pretrained with Barlow Twins self-supervised learning to reduce the need for labels. Benchmarked it against five baselines (U-Net, BT-U-Net, YOLOv8, SAM+YOLOv8, HoVerNet) and automated the full pipeline on Azure (Blob Storage, Azure ML).
Result
Segmentation 3% better than expert-annotated masks and 20% higher downstream diagnostic accuracy. Published on arXiv and under review at Pattern Recognition (Elsevier).
Let’s find out what AI can do for you
A free, no-obligation 30-minute call. You describe your processes or data, and I show you where AI could help and what a sensible first step looks like.