Artificial Intelligence for differentiation of focal liver lesions using contrast-enhanced ultrasound (CEUS)

Background and Aims: Characterization of focal liver lesions (FLL) remains an important yet challenging task in routine patient care. Contrast-enhanced ultrasound (CEUS) is a reliable tool but depends on the examiner’s expertise. Lately we showed that weakly supervised attention-based multiple instance learning (aMIL) algorithms can distinguish benign from malignant FLL. This study aims to further develop and evaluate such an algorithm.

Method: In this retrospective study, we used CEUS data from patients with four types of FLL: focal nodular hyperplasia (FNH), hemangioma, hepatocellular carcinoma (HCC), and metastasis. Features were extracted from examination video frames using a pretrained convolutional neural network. A 20% class-balanced test set was randomly removed, and the remaining data were used for training with a hyperparameter grid search and five-fold cross-validation. To choose the final set of hyperparameters we averaged the receiver operating curve (AUROC) of each five-fold crossvalidation run. The five models within this run were ensembled yielding our final classifier.

Results: Data from 370 patients (FNH n=52, Hemangioma n=149, HCC n=67, Metastasis n=102) were included. The average AUROC was 0.82 during cross-validation. The final model achieved AUROCs of 0.85 for FNH, 0.87 for hemangioma, 0.92 for HCC, and 0.93 for metastasis. Post-hoc explainability analysis suggested that the model focused on frames being judged diagnostic by human experts.

Conclusion: Our model can classify FLL according to their appearance in CEUS with very good performance without requiring resource-intensive pre-processing. After validation in prospective trials such weakly supervised algorithms could help clinicians to assess FLL.

Publication History

Article published online:
20 January 2025