Liver Nodule Anomaly Detection Using Ultra-sound Radiofrequency Signals and Variational Autoencoders.

Abstract

Detection of liver nodules in ultrasound (US) is challenging due to the low visibility in the presence of steatotic and cirrhotic livers. Anomaly detection approaches offer an unsupervised alternative by modeling "normal tissues" and identifying deviations as abnormalities. This prospective, cross-sectional study aims to evaluate the performance of a reconstruction based deep learning model for detection of liver nodules using one dimensional US radiofrequency signals (1D RF) of two-dimensional images. Adult control participants, patients suspected of metabolic dysfunction-associated steatohepatitis (MASH) characterized by biopsy, and patients with known liver nodules who underwent abdominal research US from 2018 to 2025 were included. Nodules were manually identified and segmented on B-mode US and classified using MRI or histopathology on tis sue specimens. Variationalautoencoders were trained on adjacent 1D RF lines of images from control and MASH-suspected individuals and used to reconstruct 1D RF lines for nodule detection. An anomaly score was used as a threshold for detection, and performance was assessed by the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, and specificity. The study included 177 participants (mean age ± SD: 57 years ± 12; 106 men). Mean anomaly scores were almost three times higher for patients with nodules (n = 100) than those without nodules (n = 77), providing an AUC of 0.86, an accuracy of 0.69, a sensitivity of 0.66, and a specificity of 0.97 (p < 0.001). Liver nodules can be automatically identified with variational autoencoders trained on images of individuals with normal liver or patients with suspected MASH.