Prior knowledge-guided vision-transformer-based unsupervised domain adaptation for intubation prediction in lung disease at one week

Abstract

Data-driven approaches have achieved great success in various medical image analysis tasks. However, fully-supervised data-driven approaches require unprecedentedly large amounts of labeled data and often suffer from poor generalization to unseen new data due to domain shifts. Various unsupervised domain adaptation (UDA) methods have been actively explored to solve these problems. Anatomical and spatial priors in medical imaging are common and have been incorporated into data-driven approaches to ease the need for labeled data as well as to achieve better generalization and interpretation. Inspired by the effectiveness of recent transformer-based methods in medical image analysis, the adaptability of transformer-based models has been investigated. How to incorporate prior knowledge for transformer-based UDA models remains under-explored. In this paper, we introduce a prior knowledge-guided and transformer-based unsupervised domain adaptation (PUDA) pipeline. It regularizes the vision transformer attention heads using anatomical and spatial prior information that is shared by both the source and target domain, which provides additional insight into the similarity between the underlying data distribution across domains. Besides the global alignment of class tokens, it assigns local weights to guide the token distribution alignment via adversarial training. We evaluate our proposed method on a clinical outcome prediction task, where Computed Tomography (CT) and Chest X-ray (CXR) data are collected and used to predict the intubation status of patients in a week. Abnormal lesions are regarded as anatomical and spatial prior information for this task and are annotated in the source domain scans. Extensive experiments show the effectiveness of the proposed PUDA method.