This repository contains the official implementation of ToolSeg, our proposed method for phase-informed surgical tool segmentation in Manual Small-Incision Cataract Surgery (MSICS).
Cataract surgery is one of the most common surgical procedures globally, yet most automated analysis methods have focused on Phacoemulsification, overlooking the widely used, low-cost MSICS technique. To bridge this gap, we introduce Sankara-MSICS, the first large-scale dataset for MSICS, featuring pixel-level annotations for 13 surgical tools across 18 phases in 53 surgery videos.
Our method, ToolSeg, enhances tool segmentation by:
- Leveraging surgical phase information via a novel Phase-Conditional Decoder.
- Incorporating a semi-supervised learning strategy with pseudo-labels generated by Meta’s SAM 2 foundation model.
ToolSeg achieves state-of-the-art performance, outperforming existing segmentation models on both the Sankara-MSICS and CaDIS datasets. This work aims to drive future advancements in cataract surgery automation, surgical skill analysis, and global eye care accessibility.
📢 MICCAI 2025: This work has been accepted for presentation at the International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2025.
ToolSeg is an encoder-decoder architecture with surgical phase conditioning to improve segmentation accuracy.
- Create a new conda environment and activate it.
conda create -n toolseg python=3.12
conda activate toolseg- Install the required packages in the new environment.
conda install pytorch torchvision torchaudio pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txtDownload the Sankara-MSICS dataset from here.
Make appropriate changes in config.yaml and run the following.
python main.pyAlternatively, you can override the parameters on the command line.
python main.py fold=0 condition.phase='pcd'Ensure the correct parameters are added in config.yaml.
python inference.py