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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Hepatology

Background:

• Liver nodule detection in ultrasound (US) is difficult with fatty or cirrhotic livers.
• Anomaly detection models normal tissue to find abnormalities.
• Deep learning offers an unsupervised approach for nodule identification.

Purpose of the Study:

• To evaluate a deep learning model for liver nodule detection.
• The model uses one-dimensional US radiofrequency (1D RF) signals.
• Focuses on improving detection in challenging liver conditions.

Main Methods:

• Prospective, cross-sectional study with 177 participants (controls, suspected MASH, known nodules).
• Variational autoencoders trained on 1D RF lines from normal/MASH livers.
• Anomaly scores used for nodule detection; performance measured by AUC, accuracy, sensitivity, specificity.

Main Results:

• Mean anomaly scores were significantly higher in patients with nodules (n=100) vs. without (n=77).
• Achieved an Area Under the Curve (AUC) of 0.86.
• Reported accuracy of 0.69, sensitivity of 0.66, and specificity of 0.97.

Conclusions:

• Liver nodules can be automatically identified using variational autoencoders.
• The model is trained on normal liver or suspected metabolic dysfunction-associated steatohepatitis (MASH) images.
• This deep learning approach shows promise for liver nodule detection.