UNISEG

This repository is the result of an academic project named UNISEG (Universal Segmentation) at Ecole Centrale in Nantes, France, with my classmate @damien-gautier-nantes. The aim of this project was to evaluate the performance of universal segmentation models for segmenting cancerous lesions. The UniverSeg and Segment Anything (SAM) models were tested.

They were tested using the HECKTOR dataset which regroups subjects with head and neck tumors.

For confidentiality reasons we can't expose our results in images on GitHub.

Usage :

The use of our code is described in each README of UniverSeg and SAM folders.

Download UniverSeg model

pip install git+https://github.com/JJGO/UniverSeg.git

Download SAM model

• Download a model checkpoint

  WARNING: You may need to change where the program fetches the downloaded model for SAM use.

• Then:

  pip install git+https://github.com/facebookresearch/segment-anything.git
  pip install opencv-python pycocotools matplotlib onnxruntime onnx

Clone this repository

git clone https://github.com/artastier/UNISEG.git

Improvements:

To increase the automation of the segmentation of cancerous lesions, it may be useful to develop the following pipeline:

• Automatic detection with UniverSeg model
• Remove wrong predictions from UniverSeg
• Use of SAM to load an image where we can see what UniverSeg has segmented. Prompt points on the lesions non-segmented by UniverSeg and a background point. It can be interesting to try faster version of SAM.