Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causally Guided Gaussian Perturbations for Out-Of-Distribution Generalization in Medical Imaging

Haoran Pei, Yuguang Yang, Kexin Liu, Baochang Zhang

TL;DR

The paper tackles the challenge of out-of-distribution generalization in medical imaging by introducing Causally-Guided Gaussian Perturbations (CGP), a lightweight, plug-in framework that uses Vision Transformer–derived soft causal masks to apply spatially varying Gaussian noise. By perturbing background regions more than foreground areas and weighting adversarial loss by model confidence, CGP acts as a causal intervention encouraging models to rely on invariant, causally relevant features. Evaluated on Camelyon17 (WILDS), CGP yields consistent OOD improvements over strong baselines with only a small trade-off in in-domain accuracy, and qualitative analyses show the masks focus on diagnostically meaningful regions. The approach suggests that simple, causally informed perturbations can enhance robustness and interpretability in medical imaging without heavy generative modeling or adversarial training, with potential for broader domain adoption and future refinement of mask fidelity.

Abstract

Out-of-distribution (OOD) generalization remains a central challenge in deploying deep learning models to real-world scenarios, particularly in domains such as biomedical images, where distribution shifts are both subtle and pervasive. While existing methods often pursue domain invariance through complex generative models or adversarial training, these approaches may overlook the underlying causal mechanisms of generalization.In this work, we propose Causally-Guided Gaussian Perturbations (CGP)-a lightweight framework that enhances OOD generalization by injecting spatially varying noise into input images, guided by soft causal masks derived from Vision Transformers. By applying stronger perturbations to background regions and weaker ones to foreground areas, CGP encourages the model to rely on causally relevant features rather than spurious correlations.Experimental results on the challenging WILDS benchmark Camelyon17 demonstrate consistent performance gains over state-of-the-art OOD baselines, highlighting the potential of causal perturbation as a tool for reliable and interpretable generalization.

Causally Guided Gaussian Perturbations for Out-Of-Distribution Generalization in Medical Imaging

TL;DR

The paper tackles the challenge of out-of-distribution generalization in medical imaging by introducing Causally-Guided Gaussian Perturbations (CGP), a lightweight, plug-in framework that uses Vision Transformer–derived soft causal masks to apply spatially varying Gaussian noise. By perturbing background regions more than foreground areas and weighting adversarial loss by model confidence, CGP acts as a causal intervention encouraging models to rely on invariant, causally relevant features. Evaluated on Camelyon17 (WILDS), CGP yields consistent OOD improvements over strong baselines with only a small trade-off in in-domain accuracy, and qualitative analyses show the masks focus on diagnostically meaningful regions. The approach suggests that simple, causally informed perturbations can enhance robustness and interpretability in medical imaging without heavy generative modeling or adversarial training, with potential for broader domain adoption and future refinement of mask fidelity.

Abstract

Out-of-distribution (OOD) generalization remains a central challenge in deploying deep learning models to real-world scenarios, particularly in domains such as biomedical images, where distribution shifts are both subtle and pervasive. While existing methods often pursue domain invariance through complex generative models or adversarial training, these approaches may overlook the underlying causal mechanisms of generalization.In this work, we propose Causally-Guided Gaussian Perturbations (CGP)-a lightweight framework that enhances OOD generalization by injecting spatially varying noise into input images, guided by soft causal masks derived from Vision Transformers. By applying stronger perturbations to background regions and weaker ones to foreground areas, CGP encourages the model to rely on causally relevant features rather than spurious correlations.Experimental results on the challenging WILDS benchmark Camelyon17 demonstrate consistent performance gains over state-of-the-art OOD baselines, highlighting the potential of causal perturbation as a tool for reliable and interpretable generalization.

Paper Structure

This paper contains 15 sections, 4 equations, 3 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. ViT-based Soft Causal Mask (SCM)
  5. Noise Injection Strategy
  6. Adaptive Adversarial Weighting
  7. Training Objective and Two-stage Strategy
  8. Training Workflow
  9. EXPERIMENTAL SETUP
  10. Dataset
  11. Experimental Details
  12. Hardware and Environment
  13. RESULTS AND ANALYSIS
  14. Conclusions
  15. Acknowledgment

Figures (3)

  • Figure 1: Five sample patches from the Camelyon17 dataset. Images 1–3 are from the training subset (nodes 0–2); image 4 is from the ID test subset (node 3); image 5 is from the OOD test subset (node 4).
  • Figure 2: Overview of the proposed CGP framework.
  • Figure 3: This figure shows the visualizations of representative samples from the training set, ID test set, and OOD test set (from top to bottom). Each row corresponds to one sample randomly selected from each respective dataset. The columns represent different visualization methods: (a) original image, (b) standard CAM, (c) standard Grad-CAM, (d) CAM generated by our CNN-based model, (e) Grad-CAM generated by our CNN-based model, and (f) foreground mask generated by our ViT-based model.