Experience with Single Domain Generalization in Real World Medical Imaging Deployments

TL;DR

This work tackles single-domain generalization (SDG) in real-world medical imaging deployments, focusing on rare-class detection under domain shifts across centers and devices. It proposes RareSaGe, a framework that integrates domain-invariant expert knowledge with deep learning (via vision-language models like CLIP) to identify rare classes, and DL+EKE, a discriminative system combining DL for overlap with expert knowledge for rare classes. The approach is validated on diabetic retinopathy benchmarks and demonstrated in two deployment scenarios: seizure onset zone (SOZ) detection from resting-state fMRI and coronary artery disease (CAD) detection from stress ECG, where SDG-capable baselines fall short but DL+EKE yields strong generalization (e.g., SOZ F1 ≈ 90.2% across centers; CAD unseen-test PPV ≈ 91.2%, NPV ≈ 93%). The study emphasizes that real-world deployable SDG requires iterative clinical–engineering collaboration and an explicit focus on explainability and causal shifts across domains. Future work calls for formal theories of domain causality to guide more robust SDG methods in medicine.

Abstract

A desirable property of any deployed artificial intelligence is generalization across domains, i.e. data generation distribution under a specific acquisition condition. In medical imagining applications the most coveted property for effective deployment is Single Domain Generalization (SDG), which addresses the challenge of training a model on a single domain to ensure it generalizes well to unseen target domains. In multi-center studies, differences in scanners and imaging protocols introduce domain shifts that exacerbate variability in rare class characteristics. This paper presents our experience on SDG in real life deployment for two exemplary medical imaging case studies on seizure onset zone detection using fMRI data, and stress electrocardiogram based coronary artery detection. Utilizing the commonly used application of diabetic retinopathy, we first demonstrate that state-of-the-art SDG techniques fail to achieve generalized performance across data domains. We then develop a generic expert knowledge integrated deep learning technique DL+EKE and instantiate it for the DR application and show that DL+EKE outperforms SOTA SDG methods on DR. We then deploy instances of DL+EKE technique on the two real world examples of stress ECG and resting state (rs)-fMRI and discuss issues faced with SDG techniques.

Figures (5)

• Figure 1: RareSaGe Model Schematic:1) Image Encoders: CLIP identifies the class most similar to the rare class. 2.a) DL Machine distinguishes overlap and non-overlap classes. 2.b) Knowledge Machine extracts features from the rare class, and trains a quadratic optimization model. 3) Label Predictor: Each test image is first classified by the DL classifier as either overlap or non-overlap. Non-overlapping images are further classified by the knowledge machine as rare or non-rare.
• Figure 2: ROC of DL+EKE for different knowledge override thresholds and ablation analysis of knowledge.
• Figure 3: LIME attention maps of DL+EKE focuses on ST segment for males, but on QRS and P waves for females.
• Figure 4: Expert-Guided Transformer: This 2 Encoder (E1-E2) model incorporates expert knowledge on METs and lead selection into the input stage and integrates CAD temporal characteristics into the transformer's attention layer.
• Figure 5: Use of RareSaGe on Diabetes Retinopathy grading

Theorems & Definitions (2)

• Definition 1: SDG problem definition
• Definition 2