Deep Learning Driven Evaluation of MR-guided Focused Ultrasound Ablation

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Magnetic Resonance-guided Focused Ultrasound (MRgFUS) offers incisionless thermal therapy for breast cancer.
• Current methods for assessing MRgFUS treatment efficacy using thermal and vascular MRI biomarkers lack real-time in vivo accuracy.
• Accurate, real-time assessment is crucial for effective MRgFUS treatment monitoring.

Purpose of the Study:

• To develop and validate a deep learning framework utilizing multi-parametric MRI for predicting MRgFUS treatment efficacy.
• To enable real-time, in vivo assessment of treatment outcomes immediately following MRgFUS procedures.
• To enhance the accuracy of MRgFUS efficacy evaluation beyond traditional biomarkers.

Main Methods:

• A deep learning model was developed using qualitative T1/T2-weighted images, MR temperature metrics, and quantitative T1/T2 parametric maps.
• Extensive data augmentation techniques, including principal component analysis (PCA), were employed to enhance model robustness.
• The model was trained and validated on a VX2 tumor model rabbit dataset (N=12), comparing predictions against post-treatment non-perfused volume.

Main Results:

• The deep learning biomarker achieved promising performance with traditional augmentations (Dice: 0.62, MDA: 3.7 mm).
• Incorporating PCA-based augmentation further improved boundary accuracy (Dice: 0.64, MDA: 3.0 mm).
• These results demonstrate the model's capability in predicting treatment efficacy.

Conclusions:

• A deep learning multi-parametric MRI framework shows significant potential for accurately predicting MRgFUS treatment efficacy.
• Quantitative MRI data integration enhances the accuracy and immediacy of MRgFUS outcome evaluation.
• Future work should focus on reducing multi-parametric MRI acquisition time for clinical translation.