[Paper Review] MRI-based Material Mass Density and Relative Stopping Power Estimation via Deep Learning for Proton Therapy

Yuan Gao, Chih‐Wei Chang+11 more|arXiv (Cornell University)|Oct 11, 2022

Advanced MRI Techniques and Applications4 citations

TL;DR

This study proposes a deep learning framework to estimate material mass density and relative stopping power (RSP) from MRI scans for proton therapy treatment planning. Using T1, T2, and zero echo time (ZTE) MRI sequences, the model achieves sub-1% mean absolute percentage error (MAPE) in soft tissues and below 0.3% in cortical bone substitutes, demonstrating high accuracy for MRI-only proton therapy planning.

ABSTRACT

Magnetic Resonance Imaging (MRI) is increasingly incorporated into treatment planning, because of its superior soft tissue contrast used for tumor and soft tissue delineation versus computed tomography (CT). However, MRI cannot directly provide mass density or relative stopping power (RSP) maps required for proton radiotherapy dose calculation. To demonstrate the feasibility of MRI-only based mass density and RSP estimation using deep learning (DL) for proton radiotherapy. A DL-based framework was developed to discover underlying voxel-wise correlation between MR images and mass density and RSP. Five tissue substitute phantoms including skin, muscle, adipose, 45% hydroxyapatite (HA), and spongiosa bone were customized for MRI scanning based on material composition information from ICRP reports. Two animal tissue phantoms made of pig brain and liver were prepared for DL training. In the phantom study, two DL models were trained: one containing clinical T1 and T2 MRIs and another incorporating zero echo time (ZTE) MRIs as input. In the patient application study, two DL models were trained: one including T1 and T2 MRIs as input, and one incorporating synthetic dual-energy computed tomography (sDECT) images to provide bone tissue information. In the phantom study, DL model based on T1 and T2 MRI demonstrated higher accuracy mass density and RSP estimation in skin, muscle, adipose, brain, and liver with mean absolute percentage errors (MAPE) of 0.42%, 0.14%, 0.19%, 0.78% and 0.26% for mass density and 0.30%, 0.11%, 0.16%, 0.61% and 0.23% for RSP, respectively. DL model incorporating ZTE MRI improved the accuracy of mass density and RSP estimation in 45% HA and spongiosa bone with MAPE at 0.23% and 0.09% for mass density and 0.19% and 0.07% for RSP, respectively. Results show feasibility of MRI-only based mass density and RSP estimation for proton therapy treatment planning using DL method.

Motivation & Objective

• To develop a deep learning-based method for estimating material mass density and relative stopping power (RSP) from MRI scans to enable MRI-only proton therapy treatment planning.
• To overcome the limitation of MRI in directly providing Hounsfield unit-equivalent information required for proton dose calculation.
• To evaluate the performance of deep learning models using various MRI sequences, including T1, T2, and zero echo time (ZTE) MRI, in phantoms and patient data.
• To assess the impact of incorporating synthetic dual-energy CT (sDECT) images on improving bone tissue RSP estimation accuracy.

Proposed method

• A deep learning framework was trained to learn voxel-wise correlations between multi-contrast MRI sequences (T1, T2, ZTE) and ground-truth mass density and RSP values from phantoms.
• Five tissue-equivalent phantoms (skin, muscle, adipose, 45% hydroxyapatite, spongiosa bone) and two animal tissue phantoms (pig brain, liver) were scanned to generate training data.
• Two distinct deep learning models were trained: one using only T1 and T2 MRI inputs, and another incorporating ZTE MRI or synthetic dual-energy CT (sDECT) images for improved bone tissue characterization.
• The models were validated on phantom data and applied to patient data, with performance evaluated using mean absolute percentage error (MAPE) for mass density and RSP.
• ZTE MRI was used to improve the estimation of high-density tissues such as cortical bone, where conventional T1/T2 sequences show limited contrast.
• The framework was tested in both phantom and patient studies, with sDECT used as a surrogate for high-attenuation bone structures in clinical applications.

Experimental results

Research questions

• RQ1 Can deep learning accurately estimate mass density and relative stopping power (RSP) from T1 and T2 MRI sequences alone in soft tissues and bone substitutes?
• RQ2 How does the inclusion of zero echo time (ZTE) MRI improve RSP and mass density estimation accuracy in high-density tissues like cortical bone?
• RQ3 To what extent does incorporating synthetic dual-energy CT (sDECT) images enhance the model’s ability to predict RSP in bony structures within patient MRI scans?
• RQ4 What is the achievable accuracy of deep learning-based MRI-only RSP estimation in clinical patient data compared to phantom studies?
• RQ5 Can the proposed deep learning framework achieve sub-1% mean absolute percentage error (MAPE) for mass density and RSP across diverse tissue types?

Key findings

• The deep learning model using only T1 and T2 MRI achieved a mean absolute percentage error (MAPE) of 0.42% for mass density and 0.30% for RSP in soft tissues such as skin, muscle, adipose, brain, and liver.
• Incorporating ZTE MRI reduced MAPE to 0.23% for mass density and 0.19% for RSP in 45% hydroxyapatite (HA) bone substitutes, significantly improving accuracy over T1/T2-only models.
• For spongiosa bone, the ZTE-enhanced model achieved a MAPE of 0.09% for mass density and 0.07% for RSP, demonstrating high precision in modeling trabecular bone.
• The model incorporating synthetic dual-energy CT (sDECT) images showed improved RSP estimation in patient data, particularly for cortical bone structures.
• Overall, the framework demonstrated feasibility of MRI-only proton therapy treatment planning with sub-1% MAPE across all tested tissue types, including high-density bone.