Room: Exhibit Hall | Forum 8
Purpose: The application of quantitative image analyses such as radiomics and tissue feature extraction in machine learning is often disturbed by the effect of MR intensity non-uniformity (INU). The performance of current mathematical model-based INU correction methods depends highly on how the parameters are set. This study aims to develop an automatic, deep learning based, MRI INU correction method.
Methods: The proposed method integrates residual-blocks into a 3D cycle generative adversarial network (GAN) framework to learn the mapping between uncorrected and corrected MR images. This method is capable of effectively differentiating tissue boundaries and capturing multi-slice spatial information. A cohort of 25 abdominal patients with T1-weighted MR INU image was used to evaluate the algorithm with leave-one-out cross-validation. The ground truth, corrected MRIs, were MRI with INU that was corrected by the N4ITK Bias correction filter. Imaging endpoints, including spatial non-uniformity (SNU) and intensity profile plot, were compared between the ground truth and learning-based correction results to evaluate our proposed method.
Results: The intensity profile plots of our learning-based INU corrected MR images show comparable results to the N4ITk-based INU corrected images. Furthermore, more uniform intensity distribution in the organs like the liver can be observed in the learning-based outcomes. The mean SNUs among the patient cohort computed by selecting 3 regions-of-interest in uncorrected INU MR, N4ITK corrected, and our learning-based corrected images were 0.654Â±0.167, 0.592Â±0.168, and 0.38Â±0.086, respectively.
Conclusion: We developed a novel deep-learning-based automatic MRI INU correction method with comparable results with the N4ITK corrected ground truth. Additionally, our method can outperform the ground truth in terms of organ intensity uniformity. This MRI INU correction technique could be a useful tool to accurately reduce intensity difference for quantitative MR imaging, such as, MRI-based treatment planning and monitoring tumor response.
Funding Support, Disclosures, and Conflict of Interest: NIH R01 CA215718
Not Applicable / None Entered.