Adaptive geometric-attention multi-task framework with knowledge distillation for aortic dissection detection in non-contrast CT.
Authors
Affiliations (7)
Affiliations (7)
- Guangzhou Medical University, Guangzhou Medical University (Panyu Campus), Guangzhou, 510182, CHINA.
- Department of Cardiovascular Surgery, Fujian Provincial Clinical College of Fujian Medical University, Fujian Provincial Hospital, Department of Cardiovascular Surgery, Fujian Provincial Clinical College of Fujian Medical University, Fujian Provincial Hospital, Fuzhou, 350028, CHINA.
- Shengli Clinical Medical College, Fujian Medical University, , Shengli Clinical Medical College, Fujian Medical University,, Fuzhou, 350122, CHINA.
- Department of Radiology, Union Hospital, Fujian Medical University, Department of Radiology, Union Hospital, Fujian Medical University, Fuzhou, 350122, CHINA.
- The Sixth People's Hospital of Huizhou, The Sixth People's Hospital of Huizhou, Huizhou, 516211, CHINA.
- Department of Emergency, the Second Affiliated Hospital of Guangzhou Medical University, Guangzhou, China, Guangzhou, 50000, CHINA.
- School of Medical Information Engineering, Guangzhou University of Chinese Medicine, Guangzhou, School of Medical Information Engineering, Guangzhou University of Chinese Medicine, Guangzhou, Guangzhou, 510006, CHINA.
Abstract
Aortic dissection (AD) is a life-threatening cardiovascular emergency. Non-contrast-enhanced CT (NCE-CT) could provide timely AD screening with fewer contraindications compared to contrast-enhanced CT angiography (CE-CT). However, NCE-CT examinations lack distinctive imaging characteristics of AD, leading to high rates of missed diagnoses and misdiagnoses, and increased radiologist workload. In this paper, we propose a novel end-to-end multi-task framework for automated aortic segmentation and AD detection using NCE-CT images. The framework comprises three main components: a deformable feature extractor enhancing aorta tubular-feature attention, an adaptive geometric information extraction module to optimize feature sharing between segmentation and classification tasks via the transformer cross-attention mechanism, and a knowledge distillation module transferring diagnostic information from the CE-CT-based teacher model to the NCE-CT-based student model. Multi-center tests across 3 internal and 2 external centers demonstrated that our model outperformed existing methods both for segmenting the aorta and detecting AD. Specifically, for segmenting the aorta, our framework achieved dice of 0.928 and 0.909, Jaccard index of 0.867 and 0.858, MIoU of 0.932 and 0.913, and FWIoU of 0.995 and 0.994, in internal and external testing datasets, respectively. For identifying AD patients from non-AD patients, our framework achieved accuracies of 0.911 and 0.840, sensitivities of 0.925 and 0.888, and F1-scores of 0.922 and 0.836, in internal and external testing datasets, respectively. Ablation experiment demonstrates the effectiveness of each module. The proposed model may serve as an effective diagnostic assistant for radiologists, acting as a "second pair of eyes" to assist in AD screening using NCE-CT images.