LESEN: Label-Efficient deep learning for Multi-parametric MRI-based Visual Pathway Segmentation

In recent years, multi-parametric MRI has emerged as a powerful tool in the field of visual pathway (VP) segmentation Li et al. [2021]. By combining different imaging sequences, such as T1-weighted (T1-w) and fractional anisotropy (FA), researchers have been able to overcome the limitations of individual imaging sequences and gain a more comprehensive understanding of the VP. This can be a powerful tool in the diagnosis and treatment of various diseases affecting the visual system, including optic neuritis and optic nerve glioma. During the past years, deep learning (DL) approaches, such as convolutional neural networks (CNNs) and their variants, have shown significant advancements in multi-parametric MRI-based VP segmentation Li et al. [2021], Mansoor et al. [2016], Zhao et al. [2019], Xie et al. [2023a]. These methods can automatically learn hierarchical representations from the multi-parametric MR images and effectively capture the complementary characteristics. However, one of the major challenges in these DL-based methods is the laborious and time-consuming process of obtaining labeled training data Bai et al. [2023], Yu et al. [2019]. Manual annotation is prone to errors and requires significant expertise, making it impractical for large-scale studies. Therefore, there is a critical need to develop algorithms that can achieve accurate segmentation performance even in situations with limited labeled samples.