Polyp and Surgical Instrument Segmentation with Double Encoder-Decoder Networks

TL;DR

The work tackles automated segmentation of polyps and surgical instruments in endoscopic images for the MedAI competition. It extends a previously successful double encoder-decoder framework by employing a stronger encoder, an improved optimization routine via Sharpness-Aware Minimization, and post-processing through tempered model ensembling. The authors train four rotated models and ensemble them with temperature sharpening, achieving Dice scores near 90% and showing that instrument segmentation is easier than polyp segmentation on the Kvasir-SEG and Kvasir-Instrument datasets. They discuss a precision-recall trade-off driven by temperature and note the need for approaches that handle frames without objects in test data, outlining directions for future work.

Abstract

This paper describes a solution for the MedAI competition, in which participants were required to segment both polyps and surgical instruments from endoscopic images. Our approach relies on a double encoder-decoder neural network which we have previously applied for polyp segmentation, but with a series of enhancements: a more powerful encoder architecture, an improved optimization procedure, and the post-processing of segmentations based on tempered model ensembling. Experimental results show that our method produces segmentations that show a good agreement with manual delineations provided by medical experts.

Figures (1)

• Figure 1: Graphical description of the adopted double encoder-decoder architecture for polyp and instrument segmentation.