OmniMedSeg: A Large-Scale Standardized Benchmark for Interactive Medical Image Segmentation

Medical image segmentation datasets form the foundation for developing deep learning models across diverse clinical tasks. As the field grows, the number of publicly annotated datasets has increased substantially. However, due to the sensitivity of medical data, many datasets are restricted by licenses, limited to specific challenges, or require complex access procedures. Moreover, medical imaging data is highly heterogeneous, existing in various formats that are often incompatible and difficult to use directly for model training or evaluation. In this paper, we address these issues by presenting \textit{OmniMedSeg} - a large-scale, standardized benchmark for medical image segmentation that unifies 156 openly licensed datasets spanning nine imaging modalities. We introduce a modular three-layer framework (Download, Conversion, and Data layers) that automates data retrieval, standardizes formats to PNG for 2D and NIfTI for 3D data, and preserves all original metadata with full traceability. Beyond data aggregation, we provide standardized simulation and evaluation protocols for interactive segmentation, including click, scribble, bounding box, and polygon-based robot users. ... mehrThese protocols transform any existing metric into an interactive version via AUC and Final metric computation. \textit{OmniMedSeg} is designed as a living, community-driven framework with an extensible architecture that supports continuous integration of new datasets, modalities, and protocols. This work establishes a robust foundation for fair, reproducible, and comparable benchmarking of both interactive and non-interactive segmentation methods. The converted dataset is available at: https://dx.doi.org/10.35097/qkc5m4kzran3g1yc