Generalization and Robustness of Spectral Adapters Across Imaging Domains

Determine the cross-site generalization, multi-phase performance, and robustness to protocol changes of spectral adapter-based Segment Anything Model systems for colorectal liver metastasis segmentation in computed tomography.

Background

The study evaluates DiSECT and SiGA primarily on portal-venous-phase contrast-enhanced CT scans of colorectal liver metastases. Although the dataset is multi-institutional, the authors identify unresolved questions concerning whether the reported performance transfers across sites, whether the adapters remain effective for multiphase CT, and how robust they are to variations in imaging acquisition protocols. Resolving these issues is important for assessing the clinical generalizability and deployment reliability of spectral adapter-based SAM segmentation systems.