[Value of a CT angiography-based deep learning model for differentiating SVS grade Ⅱ and grade Ⅲ blunt thoracic aortic injury: a retrospective multicenter staged study].

Abstract

Objectives: To develop a segmentation-guided multitask deep learning model for Society for Vascular Surgery (SVS) grades Ⅰ to Ⅳ classification of blunt thoracic aortic injury (BTAI) and validate its clinical utility for differentiating SVS grades Ⅱ and Ⅲ. Methods: This is a retrospective case series study. Computed tomography angiography (CTA) data were collected from 317 patients with BTAI at 12 hospitals in China, including the First Medical Center of Chinese People's Liberation Army General Hospital, West China Hospital of Sichuan University, the First Affiliated Hospital of Zhengzhou University, the Affiliated Hospital of Southwest Medical University, and other local hospitals, between January 2006 and May 2026. The cohort included 258 male and 59 female patients, with an age of (54.8±15.3) years (range:16 to 92 years). The 253-patient development cohort was divided using a patient-level random split stratified by SVS injury grade and source center at an approximate ratio of 8∶2 into a training set of 202 patients and an internal validation set of 51 patients. The 64 BTAI cases from the Affiliated Hospital of Southwest Medical University were locked in advance as an independent external test set. The model consisted of a shared encoder, a lesion-segmentation branch, and a mask-guided classification branch. The primary outcomes were the area under the receiver operating characteristic curve (AUC) and grade Ⅲ sensitivity for differentiating SVS grade Ⅱ (intramural hematoma) from grade Ⅲ (pseudoaneurysm) in the external test set. Secondary outcomes included segmentation, automated measurements, and offline reader-study results. Results: The internal validation set covered SVS grades Ⅰ to Ⅳ, with an overall grading accuracy of 92.2% (47/51) and a macro-averaged F1 score of 0.919. The external test set included 31 SVS grade Ⅱ (intramural hematoma) cases and 33 SVS grade Ⅲ (pseudoaneurysm) cases. The model had an AUC of 0.968 (95%CI:0.924 to 0.996), an accuracy of 92.2% (59/64), grade Ⅲ sensitivity of 93.9% (31/33), and grade Ⅱ specificity of 90.3% (28/31); two grade Ⅲ cases were misclassified as grade Ⅱ. The automated segmentation success rate was 96.9% (62/64), the overall lesion Dice similarity coefficient was 0.872±0.072 (range:0.46 to 0.99), and model inference time per case was (15.4±2.4) s (range:10.8 to 22.6 s). In the fixed-sequence reader study, six junior physicians had higher grade Ⅱ/Ⅲ accuracy and grade Ⅲ sensitivity and shorter interpretation times during the artificial intelligence-assisted phase than during the unaided phase. Conclusions: The segmentation-guided CTA deep learning model developed in this study performed BTAI lesion segmentation, SVS grading, and measurement of key morphologic parameters, and maintained good discrimination between SVS grades Ⅱ and Ⅲ in the independent external test set. Model assistance may help reduce missed grade Ⅲ cases by junior physicians and shorten interpretation time, thereby supporting standardized imaging assessment of BTAI.