Deep learning-based contrast-enhanced MRI using cascade networks with local supervision

Various concerns related to health, diagnostic outcomes and environmental impact have been raised recently on the wide usage of Gadolinium based contrast agents (GBCAs) in MR imaging. The purpose of this work is to propose a deep learning-based method to predict contrast enhanced MR images from the unenhanced counterpart. The proposed workflow consists of two cascade networks: the first network is trained to derive semantic features to identify the tumor regions under supervision of the tumor contours; the second network is trained to generate the synthetic contrast enhanced MR images with the input of the concatenation of the semantic features and non-contrast MR images. Qualitative and quantitative evaluations on the performance of the proposed method were conducted with MR images in the multimodal brain tumor segmentation challenge 2020 (BraTS2020) dataset. Preliminary results show that the synthetic contrast enhanced MR images were undifferentiable from the ground truth. Mean values and standard deviations of the normalized mean absolute error (NMAE), structural similarity index measurement (SSIM) and Pearson correlation coefficient (PCC) were 0.061±0.018, 0.993±0.005 and 0.996±0.005, respectively, for the whole brain; and were 0.049±0.022, 0.995±0.006 and 0.999±0.002, respectively, for the tumor regions. Utilizing cascade networks and supervising the training with tumor contours are novel for deep learning-based contrast enhanced MR image synthesis. It is expected to bypass the contrast agent usage in MR scans for diagnosis and treatment planning in radiotherapy if applied in the practice. I