Direct creation of the relative stopping power maps from MRI images using a cycleGAN deep-learning network for proton therapy

Abstract

Background

In MRI-based proton therapy, obtaining relative stopping power (RSP) maps from MRI involves converting MRI to CT, which introduces errors and uncertainties.

Purpose

We proposed a method based on deep learning to generate RSP values directly from MR images, eliminating the need for CT imaging. By bypassing the CT conversion step, our method improves the accuracy and efficiency of the RSP generation process.

Method

We implemented a cycle-consistent generative adversarial network (cycleGAN) to learn the nonlinear mapping between MR images and corresponding reference RSP maps. A total of 1000 pairs of T1-weighted MRI images from 5 patients and their related RSP maps were used to train the network. We had MR-CT paired images available, and to create reference RSP maps, we utilized the Hounsfield Look-Up Table (HLUT) to relate each Hounsfield unit (HU) value in the CT image to its corresponding RSP value.

Results

The method was evaluated with 100 random slices of head images from 5 patients. Mean absolute error (MAE), normalized mean squared error (NMSE), mutual information (MI), root mean squared error (RMSE), and mean error (ME) were used to quantify the differences between the generated and reference RSP maps. The predicted RSP maps showed an average MAE of 1.13%, average NMSE of 1.51%, average MI of 3.61, average RMSE of 1.37%, and ME of −0.95%.

Conclusion

The proposed method offers an alternative approach for deriving RSP values directly from MR images, potentially reducing costs compared to state-of-the-art approaches that rely on CT images. Our method bypasses the need for CT-based RSP calculation, providing a more streamlined and cost-effective solution.