Alexander Jaus

Alexander Jaus

Postdoctoral Researcher in Computer Vision

CVPR-W 2024

Filters, thresholds, and geodesic distances for scribble-based interactive segmentation of medical images

Z. Marinov, A. Jaus, J. Kleesiek, R. Stiefelhagen

CVPR: MedSAM Medical Image Segmentation Challenge, 2024

Paper
Interactive Segmentation Medical Imaging

Key contributions

  • A scribble-based interactive segmentation pipeline built from classical operators (thresholding, Meijering filters, geodesic distance transforms) — no large model required.
  • Outperforms MedSAM on both Dice (62.31) and Normalized Surface Dice (67.01) on the challenge validation set, at a fraction of the inference cost.
  • A modality-wise analysis (fundus, microscopy, PET, OCT) exposing where foundation models like MedSAM still fall short for thin-structure delineation.
Classical operators turn user scribbles into accurate masks across imaging modalities — faster and more accurate than MedSAM.
Classical operators turn user scribbles into accurate masks across imaging modalities — faster and more accurate than MedSAM.

How it works

Interactive segmentation plays a vital role in medical image analysis, facilitating accurate diagnosis and treatment planning through real-time interaction and rapid annotations. Scribble-based methods, where users draw over target structures, are particularly effective for delineating thin structures like vessels, providing precise pixel-level detail compared to bounding boxes. MedSAM, introduced in 2023, is optimized for bounding box inputs, which limits its effectiveness for precise interaction types such as scribbles. Additionally, it exhibits a slower inference due to its large size. To address these limitations, we evaluated simpler models such as thresholding, Meijering filters, and Geodesic Distance Transforms. These models outperformed MedSAM in segmentation accuracy and efficiency across fundus, microscopy, PET, and OCT, achieving a Dice Score of 62.31 and a Normalized Surface Dice of 67.01 on the validation set. Our findings highlight the effectiveness of traditional methods and reveal the current limitations of emerging foundation models. This comparative analysis aims to improve MedSAM’s robustness and efficiency, contributing to the development of a more reliable general model for medical image segmentation.

Citation

Marinov, Zdravko; Jaus, Alexander; Kleesiek, Jens; Stiefelhagen, Rainer; . (2024). "Filters, thresholds, and geodesic distances for scribble-based interactive segmentation of medical images." Medical Image Segmentation Challenge: 39-56.