Robustness Evaluation of a Foundation Segmentation Model Under Simulated Domain Shifts in Abdominal CT: Implications for Health Digital Twin Deployment

Basu, Sanghati

Permalink

https://hdl.handle.net/2142/133140 Copy

Description

Title

Robustness Evaluation of a Foundation Segmentation Model Under Simulated Domain Shifts in Abdominal CT: Implications for Health Digital Twin Deployment

Author(s)

Basu, Sanghati

Issue Date

2026

Keyword(s)

foundation models; Domain shift, CT segmentation; Healthcare Informatics

Date of Ingest

2026-04-24T09:17:21-05:00

Abstract

This project presents a systematic robustness evaluation of the Segment Anything Model (SAM, ViT-B) for spleen segmentation in abdominal CT imaging under simulated domain shifts. Using 1,051 non-empty slices derived from 41 volumes in the Medical Segmentation Decathlon, the study applies controlled perturbations that mimic real-world inter-scanner variability, including Gaussian noise, blur, contrast scaling, gamma correction, and resolution mismatch. The results demonstrate strong baseline performance (mean Dice = 0.9145) with consistently low failure rates and minimal degradation across all perturbation conditions (|ΔDice| < 0.01). Statistical analyses using paired Wilcoxon signed-rank tests and McNemar’s test confirm that observed variations are statistically small and not clinically significant. The findings suggest that SAM exhibits stable segmentation behavior under moderate domain shifts, supporting its role as a robust foundation baseline for medical image segmentation. From a health digital twin perspective, this work provides component-level validation for integrating foundation segmentation models into multi-stage clinical pipelines, emphasizing the importance of robustness evaluation for safe and trustworthy deployment.

Publisher

This is the author’s accepted manuscript of a paper accepted for publication in an IEEE conference. The final authenticated version will be available via IEEE Xplore.

Type of Resource

text

Genre of Resource

conference paper

Language

eng

Owning Collections