End-to-end deep learning approach for uterine artery-ovarian artery anastomosis detection from digital subtraction
Yu Lei1,2, Tao Sun1,2, Yanqiao Ren1,2
1Department of Radiology, Union Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China.
Background:
Uterine artery embolization (UAE) is a vital treatment modality for various gynecological conditions. However, complications may arise from the accidental migration of embolic agents into non-target tissues through the uterine artery-ovarian artery anastomosis (UOA). The key to addressing this issue lies in the timely detection of UOA, which is a challenging task.
Purpose:
This study aimed to develop an artificial intelligence model to detect UOA during these procedures.
Methods:
990 uterine artery angiography sequences from 610 female patients who underwent UAE were retrospectively reviewed and were divided into training, validation, internal testing, and external testing cohorts. An end-to-end artificial intelligence model called UOA Detector (UOA-D) was proposed. The detection performance was evaluated using metrics such as the area under the receiver's operating characteristic curve (AUC), average precision (AP), average precision at intersection over union (IoU) threshold 0.5 (AP_0.5), sensitivity, specificity, accuracy, and F1 score. Moreover, the impact of UOA-D on enhancing the detection capabilities of six interventional radiologists, in terms of both performance and detection time, was examined across the internal and external testing sets.
Results:
UOA-D manifested an AUC of 0.831 and 0.829 within the internal and external testing sets, with AP of 0.541 and 0.509, and AP_0.5 of 0.773 and 0.768, respectively. In the delineation of UOA-positive cases, UOA-D demonstrated sensitivity, specificity, accuracy, and F1 score of 0.978, 0.951, 0.961, and 0.978 in the internal testing set, and 0.969, 0.945, 0.959, and 0.962 in the external set, respectively. UOA-D significantly outperformed junior interventional radiologists and matched the performance of senior interventional radiologists in detecting UOA. The UOA-D-assisted strategy significantly improved the performance of junior radiologists and substantially shortened the detection time, demonstrating its potential clinical benefits.
Conclusions:
UOA-D demonstrated superior efficacy in detecting UOA through digital subtraction angiography (DSA) sequences, offering considerable advantages for clinical application.
