Detection, Localization, and Characterization of Focal Liver Lesions in Abdominal US with Deep Learning

PurposeTo train and assess the performance of a deep learning-based network designed to detect lesions in the liver parenchyma on abdominal US images, localize focal liver lesions (FLLs), and characterize them.Materials and MethodsIn this retrospective, multicenter, institutional review board-approved study, two object detectors, Faster Recurrent Convolutional Neural Network (Faster-RCNN) and DEtection vision TRansformer (DETR) were fine-tuned on a dataset of 1026 patients (n = 2551 B-mode abdominal US images between 2014 and 2018). Performances were analyzed on a test set of 48 new patients (n = 155 B-mode abdominal US images between 2018 and 2019) and were compared with that of three caregivers, one nonexpert and two experts, blinded to clinical history. A sign test was used to statistically compare accuracy, specificity, sensitivity and PPV between all raters.ResultsDETR had a specificity of 90% (95% CI: 75, 100) and a sensitivity of 97% (95% CI: 97, 97) for the detection of FLLs. Performances met or exceeded that of the three caregivers for this task. It correctly localized 80% of the lesions, and had a specificity of 81% (95% CI: 67, 91) and a sensitivity of 82% (95% CI: 62, 100) for FLL characterization (benign versus malignant) among lesions localized by all raters. Performances met or exceeded that of two experts and Faster-RCNN.ConclusionDETR demonstrated high specificity for detection, localization, and characterization of FLLs on abdominal US images.