Phase-Informed Tool Segmentation for Manual Small-Incision Cataract Surgery

TL;DR

This work addresses the lack of MSICS-focused data for automated surgical video analysis by introducing Sankara-MSICS, a large-scale dataset with 53 MSICS videos annotated for 18 surgical phases and 13 tools. It proposes ToolSeg, a phase-informed tool segmentation framework that uses a phase-conditioned decoder to leverage phase priors, combined with a MS-TCN++ phase recognizer and a SAM 2-based semi-supervised labeling pipeline that propagates masks across frames. The approach yields substantial gains in segmentation accuracy, with improvements in IoU and DSC over state-of-the-art baselines, and demonstrates strong generalization to Phaco-centric CaDIS data. These results underline the practical potential for improved training, quality control, and real-time analysis in resource-limited MSICS settings, and set the stage for joint phase recognition and tool segmentation in ophthalmic surgery.

Abstract

Cataract surgery is the most common surgical procedure globally, with a disproportionately higher burden in developing countries. While automated surgical video analysis has been explored in general surgery, its application to ophthalmic procedures remains limited. Existing works primarily focus on Phaco cataract surgery, an expensive technique not accessible in regions where cataract treatment is most needed. In contrast, Manual Small-Incision Cataract Surgery (MSICS) is the preferred low-cost, faster alternative in high-volume settings and for challenging cases. However, no dataset exists for MSICS. To address this gap, we introduce Sankara-MSICS, the first comprehensive dataset containing 53 surgical videos annotated for 18 surgical phases and 3,527 frames with 13 surgical tools at the pixel level. We benchmark this dataset on state-of-the-art models and present ToolSeg, a novel framework that enhances tool segmentation by introducing a phase-conditional decoder and a simple yet effective semi-supervised setup leveraging pseudo-labels from foundation models. Our approach significantly improves segmentation performance, achieving a $23.77\%$ to $38.10\%$ increase in mean Dice scores, with a notable boost for tools that are less prevalent and small. Furthermore, we demonstrate that ToolSeg generalizes to other surgical settings, showcasing its effectiveness on the CaDIS dataset.

Figures (10)

• Figure 1: Although the surgical tools in Images A and B appear similar, their surgical phases differ. Image A shows a Hydrodissection Cannula (from the Hydroprocedure phase), while Image B shows a Dialer (from the OVD, IOL Insertion phase). ToolSeg, our proposed method, leverages surgical phase information to accurately classify and segment instruments.
• Figure 2: Surgical phase and surgical tool co-occurence.
• Figure 3: An overview of the ToolSeg network architecture.
• Figure 4: Key components of Phase-informed Conditional Decoder layer: PAFT, DFBF, and CGate.
• Figure 5: Qualitative results of our proposed ToolSeg with SOTA methods. (A) Hydrodissection Cannula, (B) Hoskins Forceps and Rhexis Needle, (C) Hoskins Forceps and Keratome, (D) Hoskins Forceps and Blade, and (E) Vectis and Dialer.