MrGiovanni/SyntheticTumors

What it solves

This project addresses the scarcity of labeled medical data for training AI models to segment tumors. By generating highly realistic synthetic liver and pancreatic tumors, it allows AI models to be trained without relying solely on expensive and time-consuming manual labels from real medical images.

How it works

The project uses novel synthesis strategies to create synthetic tumors that are visually indistinguishable from real ones (passing a "Visual Turing Test"). These synthetic images are then used to train segmentation models, such as UNet and Swin-UNETR, which can then be applied to real tumor segmentation tasks with performance levels similar to those trained on real data.

Who it’s for

Medical imaging researchers and AI developers focusing on organ segmentation and cancer detection, specifically for liver and pancreatic tumors.

Highlights

• Label-Free Training: Enables training of AI models without requiring large sets of manually labeled real tumors.
• High Realism: Synthetic tumors are designed to be visually similar to real tumors, challenging even human observers.
• Model Compatibility: Provides pre-trained models and training scripts for multiple architectures including UNet and various sizes of Swin-UNETR (Base, Small, Tiny).
• Cross-Organ Application: Demonstrated effectiveness for both liver and pancreatic cancer detection and localization.