Using Unsupervised Domain Adaptation Semantic Segmentation for Pulmonary Embolism Detection in Computed Tomography Pulmonary Angiogram (CTPA) Images

A research team has developed a breakthrough AI system that dramatically improves pulmonary embolism detection across different medical institutions. The framework, created by Wen-Liang Lin and Yun-Chien Cheng, addresses the critical 'domain shift' problem where AI models trained at one hospital perform poorly at another due to variations in imaging equipment and protocols.

The system employs an unsupervised domain adaptation (UDA) approach using a Transformer backbone with three specialized modules: Prototype Alignment to reduce category-level distribution discrepancies, Global and Local Contrastive Learning to capture pixel-level relationships, and an Attention-based Auxiliary Local Prediction module that automatically identifies high-information slices for detecting small PE lesions. In cross-center validation between FUMPE and CAD-PE datasets, the method increased Intersection over Union (IoU) scores from 0.1152 to 0.4153 (a 260% improvement) and from 0.1705 to 0.4302 respectively.

Most impressively, the framework achieved a 69.9% Dice score in CT-to-MRI cross-modality testing on the MMWHS dataset without using any target-domain labels for model selection. This demonstrates exceptional robustness for diverse clinical environments. The approach solves two major barriers in medical AI deployment: the prohibitive cost of expert annotations for each new hospital system, and performance degradation when models encounter unfamiliar imaging data.