Deep Learning Driven Evaluation of MR-guided Focused Ultrasound Ablation.
Authors
Abstract
Magnetic resonance-guided focused ultrasound (MRgFUS) thermal therapy is a promising incisionless procedure for breast cancer treatment. for assessing treatment efficacy. Current approaches rely on thermal and vascular MRI-derived biomarkers to assess treatment efficacy. However, these techniques are not sufficiently accurate for real-time in vivo assessment. To address this challenge, we propose a deep-learning framework that utilizes multi-parametric MRI to predict treatment efficacy at the time of treatment. The presented model utilizes qualitative T1- and T2-weighted images, MR temperature-derived metrics and quantitative T1 and T2 parametric maps to predict treatment efficacy. Model robustness was enhanced via extensive data augmentation techniques. To train and validate this approach, an MRgFUS ablation study was conducted on a dataset of VX2 tumor model rabbits (N=12). The predictive power of the multi-parametric MRI model was evaluated by comparing the trained model's predictions against the three-day post-treatment non-perfused volume. The deep learning based biomarker trained with traditional augmentations (random rigid rotations, random cropping, label shifts and Gaussian noise) showed promising performance (Dice: 0.62, MDA: 3.7 mm). Adding principal component analysis based augmentation further refined boundary accuracy (Dice: 0.64, MDA: 3.0 mm). The results highlight the potential of a deep learning multi-parametric MRI framework for accurately predicting MRgFUS treatment efficacy. While improvements would be required to reduce the acquisition time of the multi-parametric MRI protocol, inclusion of quantitative MRI data offers a pathway toward more accurate and immediate evaluation of MRgFUS therapy outcomes.