Learning Multiparametric Biomarkers for Assessing MR-Guided Focused Ultrasound Treatment of Malignant Tumors

TL;DR

This study introduces a noncontrast, deep learning-based multiparametric MRI biomarker for real-time assessment of MR-guided focused ultrasound treatment efficacy in tumors, outperforming current contrast-enhanced methods.

Contribution

The paper presents a novel deep learning approach using noncontrast MRI data to predict treatment outcomes during MRgFUS, eliminating the need for contrast agents and improving accuracy.

Findings

Predicted follow-up nonperfused volume with DICE of 0.71

Outperformed current clinical standard with DICE of 0.53

Validated with voxel-wise correlation using a novel registration algorithm

Abstract

Noninvasive MR-guided focused ultrasound (MRgFUS) treatments are promising alternatives to the surgical removal of malignant tumors. A significant challenge is assessing the viability of treated tissue during and immediately after MRgFUS procedures. Current clinical assessment uses the nonperfused volume (NPV) biomarker immediately after treatment from contrast-enhanced MRI. The NPV has variable accuracy, and the use of contrast agent prevents continuing MRgFUS treatment if tumor coverage is inadequate. This work presents a novel, noncontrast, learned multiparametric MR biomarker that can be used during treatment for intratreatment assessment, validated in a VX2 rabbit tumor model. A deep convolutional neural network was trained on noncontrast multiparametric MR images using the NPV biomarker from follow-up MR imaging (3-5 days after MRgFUS treatment) as the accurate label of nonviable tissue. A novel volume-conserving registration algorithm yielded a voxel-wise correlation between treatment and follow-up NPV, providing a rigorous validation of the biomarker. The learned noncontrast multiparametric MR biomarker predicted the follow-up NPV with an average DICE coefficient of 0.71, substantially outperforming the current clinical standard (DICE coefficient = 0.53). Noncontrast multiparametric MR imaging integrated with a deep convolutional neural network provides a more accurate prediction of MRgFUS treatment outcome than current contrast-based techniques.