Domain knowledge based comprehensive segmentation of Type-A aortic dissection with clinically-oriented evaluation

Abstract

Type-A aortic dissection (TAAD) is a cardiac emergency in which rapid diagnosis, prognosis prediction, and surgical planning are critical for patient survival. A comprehensive understanding of the anatomic structures and related features of TAAD patients is the key to completing these tasks. However, due to the emergent nature of this disease and requirement of advanced expertise, manual segmentation of these anatomic structures is not routinely available in clinical practice. Currently, automatic segmentation of TAAD is a focus of the cardiovascular imaging research. However, existing works have two limitations: no comprehensive public dataset and lack of clinically-oriented evaluation. To address these limitations, in this paper we propose imageTAAD, the first comprehensive segmentation dataset of TAAD with clinically-oriented evaluation. The dataset is comprised of 120 cases, and each case is annotated by medical experts with 35 foreground classes reflecting the clinical needs for diagnosis, prognosis prediction and surgical planning for TAAD. In addition, we have identified four key clinical features for clinically-oriented evaluation. We also propose SegTAAD, a baseline method for comprehensive segmentation of TAAD. SegTAAD utilizes two pieces of domain knowledge: (1) the boundaries play a key role in the evaluation of clinical features, and can enhance the segmentation performance, and (2) the tear is located between TL and FL. We have conducted intensive experiments with a variety of state-of-the-art (SOTA) methods, and experimental results have shown that our method achieves SOTA performance on the ImageTAAD dataset in terms of overall DSC score, 95% Hausdorff distance, and four clinical features. In our study, we also found an interesting phenomenon that a higher DSC score does not necessarily indicate better accuracy in clinical feature extraction. All the dataset, code and trained models have been published (Xiaowei, 2024).