Objectives: The aim of this study was to investigate the detection efficacy of deep learning (DL) for automatic breast ultrasound (ABUS) and factors affecting its efficacy. Methods: Women who underwent ABUS and handheld ultrasound from May 2016 to June 2017 (N = 397) were enrolled and divided into training (n = 163 patients with breast cancer and 33 with benign lesions), test (n = 57) and control (n = 144) groups. A convolutional neural network was optimised to detect lesions in ABUS. The sensitivity and false positives (FPs) were evaluated and compared for different breast tissue compositions, lesion sizes, morphologies and echo patterns. Results: In the training set, with 688 lesion regions (LRs), the network achieved sensitivities of 93.8%, 97.2 and 100%, based on volume, lesion and patient, respectively, with 1.9 FPs per volume. In the test group with 247 LRs, the sensitivities were 92.7%, 94.5 and 96.5%, respectively, with 2.4 FPs per volume. The control group, with 900 volumes, showed 0.24 FPs per volume. The sensitivity was 98% for lesions > 1 cm3, but 87% for those ≤1 cm3 (p < 0.05). Similar sensitivities and FPs were observed for different breast tissue compositions (homogeneous, 97.5%, 2.1; heterogeneous, 93.6%, 2.1), lesion morphologies (mass, 96.3%, 2.1; non-mass, 95.8%, 2.0) and echo patterns (homogeneous, 96.1%, 2.1; heterogeneous 96.8%, 2.1). Conclusions: DL had high detection sensitivity with a low FP but was affected by lesion size. Advances in knowledge: DL is technically feasible for the automatic detection of lesions in ABUS.
Deep learning-based method to accurately estimate breast tissue optical properties in the presence of the chest wall