Machine learning-driven ultrasound echo feature analysis for accurate classification and area prediction of HIFU-induced lesions: ex vivo study.

In high-intensity focused ultrasound (HIFU) treatment, precise and reliable prediction of lesion phenotype and area yet remains challenging due to the intricate structures and heterogeneous responses of biological tissues. Conventional monitoring technologies such as B-mode sonography cannot distinguish among the various HIFU-induced lesion patterns or quantify their extent. To address this limitation, we systematically varied acoustic parameters-duty cycle, pulse duration, sonication time, and acoustic power-to create five distinct lesion phenotypes in ex vivo bovine livers. Ultrasonic echo signals from the HIFU focal region were acquired before and after HIFU treatment, from which 21 features were extracted and fed to three machine learning (ML) models: random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM). On the augmented dataset, XGBoost outperformed the other models, achieving an average classification accuracy of 82.6% and an area under the receiver-operating-characteristic curve (AUC) exceeding 0.93 for every lesion category. The same model also demonstrated excellent predictive performance for lesion areas, with R2 > 0.84 for three of the five phenotypes. Subsequent feature-importance analysis revealed that each lesion phenotype exhibits a unique time-frequency signature, providing discriminative information for robust lesion classification. This ex vivo study demonstrated ultrasound echo-feature analysis coupled with ML can simultaneously identify HIFU lesion type and estimate lesion area, thereby providing a methodological basis for future intra-procedural real-time HIFU monitoring and underscoring the potential of echo features for feedback-controlled treatment.