Cataract surgery is the primary treatment for cataracts and requires precise surgical skills that are critical for successful outcomes. To enhance the training of new surgeons, automated analysis of surgical videos can provide valuable feedback, particularly in identifying instruments and anatomical structures. However, current segmentation methods often fail to accurately identify these components with consistent performance, which limits their effectiveness as training tools. To address this, we propose CatSeg, an advanced surgical scene segmentation model designed to improve the accuracy of both instrument and anatomical segmentation during cataract surgery. CatSeg employs a pyramidal convolution backbone and a Multi-Scale Instrument Feature Attention (MSIFA) module, enhancing edge feature preservation for better instrument identification. Evaluated on the Cataract-1k and CaDIS benchmarks, CatSeg achieved an improved mean IoU of +8.13% (0.8107 → 0.8920) and +1.35% (0.8670 → 0.8805), respectively. With an efficient operation speed of 6.35 to 7.66 frames per second, CatSeg is well-suited for surgical video analysis, paving the way for enhanced training and analysis in the development of new surgeons. Source code is available at: https://github.com/Muraam-Abdel-Ghani/CatSeg/ . • CatSeg is built upon pyramidal-convolution backbone with rich multiscale features. • An MSIFA module boosts instrument edge features using multi-scale attention. • CatSeg sets new SOTA on Cataract-1k and CaDIS with major mIoU and Dice gains. • Ablation studies confirm the efficacy of proposed architecture. • CatSeg runs at 6.35–7.66 FPS, suitable for postop surgical video analysis.
Abdel-Ghani et al. (Sat,) studied this question.