Domain Generalization for Medical Imaging Classification with Linear-Dependency Regularization

Haoliang Li, Yufei Wang, Renjie Wan, Shiqi Wang, Tie-Qiang Li, Alex C. Kot

Abstract

Deep learning models in medical imaging often suffer from performance degradation when deployed across different clinical sites due to domain shift. This paper introduces Linear-Dependency Domain Generalization (LDDG), a framework designed to improve model robustness to unseen environments without requiring target domain data during training. By enforcing linear dependency in the latent feature space and utilizing variational encoding, LDDG captures domain-invariant features. Experimental results on skin lesion classification and spinal cord segmentation demonstrate that LDDG consistently outperforms state-of-the-art domain generalization methods.

Introduction

The deployment of AI in clinical settings is frequently hindered by "domain shift"—variations in image characteristics caused by different scanners, protocols, or patient populations. While Domain Adaptation (DA) can mitigate this by using unlabeled target data, Domain Generalization (DG) aims to build models that generalize to entirely unseen domains. This is particularly critical in medical imaging, where data privacy and proprietary constraints often prevent access to target site data during the development phase.

Problem Statement

Standard DG methods often rely on extensive data augmentation to simulate potential domain variations. However, in medical imaging, domain shifts are often structured and relatively compact (e.g., subtle contrast changes in MRI). Manually designing augmentations that accurately reflect these clinical realities is challenging. There is a need for a more principled approach that learns the underlying structure of domain variability directly from the data.

Proposed Method: Linear-Dependency Domain Generalization (LDDG)

The LDDG framework focuses on regularizing the latent feature space to ensure that learned representations are robust across domains. It employs two primary mechanisms:

  1. Linear-Dependency (Rank) Regularization: Based on the assumption that features for a specific category across different domains should be linearly dependent, the authors constrain the rank of the feature matrix to equal the number of categories ($C$). This forces the model to discard domain-specific noise and focus on shared, class-relevant information.
  2. Variational Encoding: To prevent the formation of domain-specific clusters in the latent space, a Kullback-Leibler (KL) divergence loss is used to align the feature distributions with a standard Gaussian prior.

Theoretical analysis provided by the authors shows that these constraints effectively minimize the upper bound of the generalization error on unseen target domains.

Experiments and Results

The LDDG framework was evaluated on two distinct medical imaging tasks: skin lesion classification and spinal cord gray matter segmentation.

Skin Lesion Classification

Using seven public datasets (e.g., HAM10000, MSK, PH2), the model was trained in a "leave-one-out" fashion, where one dataset served as the unseen target. LDDG achieved an average accuracy of 66.97%, outperforming the "DeepAll" baseline (62.87%) and other DG methods like MASF and MLDG.

Quantitative results for skin lesion classification across various target datasets, showing LDDG (Ours) consistently achieving the highest average accuracy.

Spinal Cord Segmentation

The method was also tested on multi-site MRI data for gray matter segmentation. LDDG achieved a Dice Similarity Coefficient (DSC) of 0.7881, a significant improvement over the baseline 0.7425. Qualitative analysis shows that LDDG produces cleaner segmentation boundaries compared to other methods, which often struggle with domain-specific artifacts.

Qualitative comparison of spinal cord gray matter segmentation across four different sites. LDDG (Ours) maintains high fidelity to the ground truth (gt) compared to other baselines.

Conclusion

LDDG offers a robust solution for domain generalization in medical imaging by leveraging the structured nature of clinical domain shifts. By regularizing the latent space through rank constraints and variational encoding, the model learns features that are inherently more shareable across different medical environments. Future work may explore non-linear dependencies for more complex natural image domains and address the computational overhead associated with rank regularization.

Summary updated

Abstract

Recently, we have witnessed great progress in the field of medical imaging classification by adopting deep neural networks. However, the recent advanced models still require accessing sufficiently large and representative datasets for training, which is often unfeasible in clinically realistic environments. When trained on limited datasets, the deep neural network is lack of generalization capability, as the trained deep neural network on data within a certain distribution (e.g. the data captured by a certain device vendor or patient population) may not be able to generalize to the data with another distribution. In this paper, we introduce a simple but effective approach to improve the generalization capability of deep neural networks in the field of medical imaging classification. Motivated by the observation that the domain variability of the medical images is to some extent compact, we propose to learn a representative feature space through variational encoding with a novel linear-dependency regularization term to capture the shareable information among medical data collected from different domains. As a result, the trained neural network is expected to equip with better generalization capability to the "unseen" medical data. Experimental results on two challenging medical imaging classification tasks indicate that our method can achieve better cross-domain generalization capability compared with state-of-the-art baselines.
BibTeX
@inproceedings{li2020domain,
  author = {Haoliang Li and Yufei Wang and Renjie Wan and Shiqi Wang and Tie-Qiang Li and Alex C. Kot},
  title = {{Domain Generalization for Medical Imaging Classification with Linear-Dependency Regularization}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
}
Figure