Weakly Supervised Deep Learning Can Analyze Focal Liver Lesions in Contrast-Enhanced Ultrasound

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Assessing malignancy of focal liver lesions (FLLs) is crucial but challenging.
• Contrast-enhanced ultrasound (CEUS) is a reliable tool, but examiner expertise is key.
• Artificial intelligence (AI) offers potential for diagnostic assistance.

Purpose of the Study:

• To evaluate a weakly supervised deep learning model for classifying FLLs as malignant or benign.
• To assess the model's performance in distinguishing between malignant and benign liver tumors using CEUS data.

Main Methods:

• Retrospective feasibility study using CEUS data from 370 patients.
• Trained a weakly supervised attention-based multiple instance learning algorithm with 5-fold cross-validation.
• Tested the aggregated model on a held-out dataset and used explainability methods.

Main Results:

• The model achieved a mean area under the receiver operating curve of 0.94 on the test set.
• Performance metrics included 80.0% accuracy, 81.8% sensitivity, and 84.6% specificity.
• Explainability analysis indicated the model focuses on clinically relevant image features.

Conclusions:

• Weakly supervised deep learning effectively classifies malignancy in CEUS for FLLs.
• This AI approach has the potential to support clinical decision-making in routine practice.