Can deep learning be used to produce high quality, synthetic, post-contrast MRI brain images without the use of gadolinium-based contrast agents?
According to a Thursday presentation by Gowtham Murugesan, PhD, Department of Radiology, University of Texas Southwestern Medical Center, Dallas, no dose gadolinium contrast using deep learning is feasible.
Gadolinium shortens the spin-lattice relaxation (T1) in tissues where the agent accumulates, resulting in brighter signal in these tissues on post-contrast T1-weighted images. In addition, gadolinium increases tissue contrast by accentuating areas where contrast agents have leaked through the blood-brain barrier.
However, evidence from research has also shown that gadolinium can accumulate in various non-target tissues (including the brain, bone and kidneys), and in 2018 the U.S. Federal Drug Administration issued a warning regarding gadolinium retention.
"This has introduced a lot of interest in identifying methods to develop contrast without injecting gadolinium," Dr. Murugesan explained.
For example, in this study researchers sought to demonstrate the feasibility of deep learning models and methods to generate T1 post-contrast images using non-contrast MRI images in primary brain tumor patients. The model used a 10% gadolinium dose to create contrast-enhanced brain images that could be used to predict full-dose images.
The dataset included 335 subjects from the Multimodal Brain Tumor Segmentation (BraTS) Challenge 2019 who were used for training the model, while a set of 125 subjects from BraTS 2019 was used as test data.
The model was quantitatively evaluated by calculating the peak signal-to-noise ratio (PSNR), normalized mean squared error (NMSE), and structural similarity index (SSIM).
The average PSNR, NMSE and SSIM for the whole brain was 64.35, 0.03, and 0.91, while the whole tumor and enhancing tumor regions demonstrated lower SSIM and PSNR, Dr. Murugesan reported. "The Dice coefficients for enhancing tumor on 125 validation subjects was 0.32, 0.35 and 0.62 for uncorrected (whole brain), corrected for whole tumor, and corrected for ET, respectively," he said.
Qualitative evaluation was performed by two radiologists who independently scored the synthesized post-contrast images using a 3-point scale (1=poor, 2=good, 3=excellent) based on image quality and synthesized contrast enhancement. According to the consensus ratings of the two radiologists, the model was able to synthesize contrast enhancement with excellent, good, and poor results in 49, 61, and 15 cases, respectively.
"The model failed to predict gadolinium enhancement in subjects where one or more of the input sequences (T1W, T2W, or FLAIR) had significant motion artifact, or where the tumor was isointense to normal brain parenchyma on both T2W and FLAIR sequences," Dr. Murugesan explained. "The failure of the model to predict enhancement in the latter scenario may be due to inadequate representation of tumors with these imaging features in the training dataset. This could be alleviated through the incorporation of additional larger datasets for training in the future. Distortion of image quality due to image artifacts, such as motion, could be separately addressed either by preventing them or correcting them retrospectively."
The proposed deep learning model is not ready for clinical translation, according to Dr. Murugesan. However, he added, the study demonstrates the feasibility of using deep learning methods to synthesize no dose gadolinium.
"Larger patient datasets for training, testing and validation are needed to ensure the clinical viability of this novel approach," Dr. Murugesan said "Additionally, this approach could also be developed and applied for other neurological diseases in addition to primary brain tumors."
For More Information:
View the RSNA 2020 session No Dose Gadolinium Contrast Using Deep Learning — SSNR15-06 at RSNA2020.RSNA.org.