270 MS T1/T2 pairs acquired at 3T were used to train the algorithm. Skulls were removed and each pair was co-registered for better pixel to pixel correspondence. We simultaneously synthesized 3D FLAIR images and lesion segmentation masks using a multi-modal supervised deep learning approach. Its architecture entails a common U-NET as an encoder, an average function for combining latent representation and a connected decoder for each output. Our multi-modal approach benefits from multiple inputs i.e. T1, T2 and Lesion Probability Map (LPM). For validation purposes, three comparative methods were implemented; Mean Absolute Error (MAE, best=0), Peak Signal-to-Noise Ratio (PSNR, best=50dB) and Structural Similarity index (SSIM, best=1). Comparisons were done between synthetic FLAIR, segmented lesions as output and the acquired FLAIR, manual segmentation as ground truth.

63 additional MS brain pairs were then used to validate the trained algorithm. MAE, PSNR and SSIM are computed for each slice and averaged altogether. MAE for synthetic FLAIR/lesion segmentation had a value of 0.087/0.0013 as the average absolute difference between two images; average PSNR value was 31.32/29.46 for the peak signal to noise ratio; SSIM between two images were 0.9/0.99, respectively.

Synthetic images and automated lesion segmentation may represent a breakthrough for quantitative analysis of heterogeneous MS scans. Our method is fully automated and shows excellent agreement in lesion segmentation against a ground truth. Adding a lesion probability map as additional input resulted in increase accuracy and faster convergence of our algorithm.