Unsupervised learning methods for magnetization transfer (MT) quantification and lesion segmentation in Brain MRI

Abstract

Deep learning is a powerful data-driven approach based on artificial neural networks with representation learning. Especially, supervised learning approaches have shown outstanding performances in the several applications of industrials and researches. In medical image fields, supervised learning shows great potentials in image reconstruction and lesion segmentation. However, as a common issue in supervised learning, the good quality of label data should be prepared as ground-truth to make a neural network achieve high performance. In medical image fields where collecting ground-truth labels is difficult or even impossible, supervised learning approaches might have limited performances and be not applicable to real clinical practices.

Tissue quantification and lesion segmentation are two typical tasks of image processing in magnetic resonance imaging (MRI). In this study, unsupervised deep learning algorithms for the two tasks are proposed, which do not use ground-truth information in training phases but provide acceptable performances. In the case of the quantification of tissue characteristics, simulated training data are generated by modeling MR signals and are used to train a neural network. Test experiments with in-vivo MR images demonstrated that the generated simulation data could reflect actual in-vivo environments and could play a successful role as training data. The results show that the proposed approach is more accurate and higher computational efficiency than conventional quantification approaches. In the case of lesion segmentation, an unsupervised algorithm is proposed by using multi-contrast MRI information. By learning the normal tissue characteristics of MRI, anomalies distinct from normal tissues are detected as lesion area. Experimental results showed that the proposed method could perform lesion segmentation without pixel-level annotation labels. Also, the proposed method could provide higher accuracy than the state-of-the-art approaches. Details are explained in the main manuscript.