Taking a Step Back: Revisiting Classical Approaches for Efficient Interactive Segmentation of Medical Images
Abstract
Interactive segmentation plays a pivotal role in medical image analysis for several reasons. It enables clinicians to precisely delineate regions of interest for accurate diagnosis and treatment planning while also allowing for real-time interaction with rapid annotations without workflow interruptions. While the emergence of MedSAM in 2023 presented a promising solution with its modality-agnostic model, its efficiency is hindered by its large size, resulting in long inference times. In response, we revisited simpler models such as thresholding, k-means clustering, and shape-based slice interpolation for efficient interactive segmentation tailored to specific modalities. Surprisingly, these rudimentary expert models outperformed MedSAM in terms of both segmentation performance and computational efficiency on multiple imaging modalities reaching a Dice score of 85.65 and a Normalized Surface Dice of 86.68 on the validation set. Our findings show the need to compare to older, simpler approaches to unveil the limitations of emerging foundation models. By examining these approaches, we aim to discover why MedSAM fails on certain modalities and enhance its robustness and efficiency leading to a more reliable general model for the segmentation of medical images.
Jens Kleesiek
Professor of Translational Image-guided Oncology
