Automated Tumor and FUS Lesion Quantification on Multi-Frequency Harmonic Motion and B-Mode Imaging Using a Multi-Modality Neural Network

Accurate tumor and FUS lesion quantification is critical for characterizing breast tumors and detecting lesions during focused ultrasound (FUS) ablation therapy. Harmonic Motion Imaging (HMI) is an ultrasound elasticity imaging that measures the mechanical properties of tissue using amplitude-modulated acoustic radiation force (AM-ARF). By estimating on-axis oscillatory displacements, HMI-derived displacement images reflect local tissue stiffness. Multi-frequency HMI (MF-HMI) advances this technique by measuring displacements at various AM frequencies simultaneously. However, accurately estimating inclusion size requires additional work due to inconsistencies in perceived sizes across AM frequencies, resulting from different boundary effects. This study aims to develop an automated tumor and FUS lesion quantification method using HMINet, a segmentation neural network on MF-HMI and B-mode images. Its performance was tested on phantom inclusions, in vivo murine tumors, an in vivo human breast tumor, and a FUS lesion, with average Dice Similarity Scores of 0.89, 0.82, 0.79, and 0.87, respectively. The main contribution lies in leveraging diverse MF-HMI information and integrating different imaging modalities (HMI and B-mode) for accurate boundary detection. It facilitates the automated tumor response prediction to neoadjuvant chemotherapy (NACT) and the real-time application of HMI-guided focused ultrasound ablation therapy (HMIgFUS).