AI generated synthetic STIR of the lumbar spine from T1 and T2 MRI sequences trained with open-source algorithms

BACKGROUND AND PURPOSE: To train and evaluate an open-source generative adversarial networks (GANs) to create synthetic lumbar spine MRI STIR volumes from T1 and T2 sequences, providing a proof-of-concept that could allow for faster MRI examinations.

MATERIALS AND METHODS: 1817 MRI examinations with sagittal T1, T2, and STIR sequences were accumulated and randomly divided into training, validation, and test sets. GANs were trained to create synthetic STIR volumes using the T1 and T2 volumes as inputs, optimized using the validation set, then applied to the test set. Acquired and synthetic test set volumes were independently evaluated in a blinded, randomized fashion by three radiologists specializing in musculoskeletal imaging and neuroradiology. Readers assessed image quality, motion artifacts, perceived likelihood of the volume being acquired or synthetic, and presence of 7 pathologies.

RESULTS: The optimal model leveraged a customized loss function that accentuated foreground pixels, achieving a structural similarity imaging metric (SSIM) of 0.842, mean absolute error (MAE) of 0.028, and peak signal to noise ratio (PSNR) of 26.367. Radiologists could distinguish synthetic from acquired volumes; however, the synthetic volumes were of equal or better quality in 77% of test patients and demonstrated equivalent or decreased motion artifacts in 78% of test patients. For common pathologies, the synthetic volumes had high positive predictive value (75-100%) but lower sensitivity (0-67%).

CONCLUSIONS: This work links objective computer vision performance metrics and subject clinical evaluation of synthetic spine MRIs using open-source and reproducible methodologies. High-quality synthetic volumes are generated, reproducing many important pathologies, demonstrating a potential means for expediting imaging protocols.

ABBREVIATIONS: AI = Artificial Intelligence; GANs = general adversarial networks; aqSTIR = acquired STIR volume; sSTIR = synthetically generated STIR volume; SSIM = structural similarity imaging metric; PSNR = peak signal to noise ratio; MAE = mean absolute error.