Detecting Performance Degradation under Data Shift in Pathology Vision-Language Model

TL;DR

This work investigates performance degradation of pathology vision-language models under real-world data shift and develops a practical monitoring framework. It analyzes both input-level data shift and output-level prediction signals, introducing DomainSAT for systematic shift analysis and visualization, and a label-free Confidence-based Degradation Indicator (CDI) to directly track degradation in model confidence. The study shows that while input-shift detection provides early warning of distribution changes, CDI closely tracks degradation and offers a complementary signal; together they enable more reliable monitoring, as demonstrated on a large Camelyon17-based dataset with scanner-induced variability. The proposed DomainSAT-CDI framework offers actionable tools for monitoring the reliability of digital pathology models in deployment, with open-source availability to facilitate adoption and extension.

Abstract

Vision-Language Models have demonstrated strong potential in medical image analysis and disease diagnosis. However, after deployment, their performance may deteriorate when the input data distribution shifts from that observed during development. Detecting such performance degradation is essential for clinical reliability, yet remains challenging for large pre-trained VLMs operating without labeled data. In this study, we investigate performance degradation detection under data shift in a state-of-the-art pathology VLM. We examine both input-level data shift and output-level prediction behavior to understand their respective roles in monitoring model reliability. To facilitate systematic analysis of input data shift, we develop DomainSAT, a lightweight toolbox with a graphical interface that integrates representative shift detection algorithms and enables intuitive exploration of data shift. Our analysis shows that while input data shift detection is effective at identifying distributional changes and providing early diagnostic signals, it does not always correspond to actual performance degradation. Motivated by this observation, we further study output-based monitoring and introduce a label-free, confidence-based degradation indicator that directly captures changes in model prediction confidence. We find that this indicator exhibits a close relationship with performance degradation and serves as an effective complement to input shift detection. Experiments on a large-scale pathology dataset for tumor classification demonstrate that combining input data shift detection and output confidence-based indicators enables more reliable detection and interpretation of performance degradation in VLMs under data shift. These findings provide a practical and complementary framework for monitoring the reliability of foundation models in digital pathology.

Figures (7)

• Figure 1: Overview of domain (data) shift across different medical sites and its impact on performance of medical AI models. An AI model deployed in Site A operates well, but when applied to Site B, the model suffers degradation under data shift.
• Figure 2: Overview of the DomainSAT workflow. The toolbox comprises three main components: (1) a data loading module for importing and preprocessing datasets; (2) an algorithm module that integrates multiple domain shift detection methods; and (3) an output module that generates detection results and visualizations.
• Figure 3: Illustration of the pretrained vision-language model (PathGen-CLIP) applied to histopathology image classification. The inputs are a histopathology image and two text prompts ("An H&E image of a tumor patch" and "An H&E image of a normal patch"). The image and text are separately encoded and compared via similarity scoring. The resulting similarity scores are normalized into predicted probabilities for the two classes (tumor vs. normal).
• Figure 4: Input data shift and performance degradation of a state-of-the-art pathology VLM on out-of-distribution (OOD) sites. (a) Data shift score (mean ± std) for 20 OOD subgroups (batches) from each OOD site, measured relative to the baseline of in-distribution (ID) dataset. (b) Performance degradation (AUC drop) of the VLM across the 20 OOD batches; the baseline is the model's tumor-classification AUC on the full ID site.
• Figure 5: Predicted tumor probability distributions on the in-distribution (ID) site and two out-of-distribution (OOD) sites. Each panel shows histograms of predicted tumor probabilities for true normal (blue) and true tumor (orange) samples. On OOD Site 1, predictions collapse toward the decision boundary ($0.5$). On OOD Site 2, predictions are more separated.