Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Alou Diakite, Cheng Li, Lei Xie, Yuanjing Feng, Hua Han, Shanshan Wang

TL;DR

LESEN tackles the label scarcity problem in multi-parametric MRI visual pathway segmentation by introducing a label-efficient self-ensembling mean teacher framework that leverages both labeled and unlabeled data. The architecture uses a two-subnetwork (student and teacher) setup with a U-Net backbone and spatial attention to fuse T1-weighted and FA modalities, guided by a supervised loss $L_{sup}$ and an unsupervised consistency loss $L_{cons}$, where the teacher EMA $ heta^{T}$ guides the student via $ heta_{k}^{T} = abla heta_{k-1}^{T} + (1 - abla) heta_{k}^{S}$ and unlabeled samples are selected by RUSS with $ q = int(b \times e^{-(1-\text{epoch}/\text{total\_epoch})^2})$. On the HCP dataset, LESEN achieves a Dice score of $83.6\%$ and an ASD of $0.21\,\text{mm}$, outperforming both supervised and other semi-supervised methods, demonstrating strong label-efficient performance for VP segmentation. The authors provide code at the project URL to facilitate adoption in clinical and research VP analysis, underscoring the practical impact of leveraging unlabeled data in multimodal MRI analysis.

Abstract

Recent research has shown the potential of deep learning in multi-parametric MRI-based visual pathway (VP) segmentation. However, obtaining labeled data for training is laborious and time-consuming. Therefore, it is crucial to develop effective algorithms in situations with limited labeled samples. In this work, we propose a label-efficient deep learning method with self-ensembling (LESEN). LESEN incorporates supervised and unsupervised losses, enabling the student and teacher models to mutually learn from each other, forming a self-ensembling mean teacher framework. Additionally, we introduce a reliable unlabeled sample selection (RUSS) mechanism to further enhance LESEN's effectiveness. Our experiments on the human connectome project (HCP) dataset demonstrate the superior performance of our method when compared to state-of-the-art techniques, advancing multimodal VP segmentation for comprehensive analysis in clinical and research settings. The implementation code will be available at: https://github.com/aldiak/Semi-Supervised-Multimodal-Visual-Pathway- Delineation.

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

TL;DR

LESEN tackles the label scarcity problem in multi-parametric MRI visual pathway segmentation by introducing a label-efficient self-ensembling mean teacher framework that leverages both labeled and unlabeled data. The architecture uses a two-subnetwork (student and teacher) setup with a U-Net backbone and spatial attention to fuse T1-weighted and FA modalities, guided by a supervised loss and an unsupervised consistency loss , where the teacher EMA guides the student via and unlabeled samples are selected by RUSS with . On the HCP dataset, LESEN achieves a Dice score of and an ASD of , outperforming both supervised and other semi-supervised methods, demonstrating strong label-efficient performance for VP segmentation. The authors provide code at the project URL to facilitate adoption in clinical and research VP analysis, underscoring the practical impact of leveraging unlabeled data in multimodal MRI analysis.

Abstract

Recent research has shown the potential of deep learning in multi-parametric MRI-based visual pathway (VP) segmentation. However, obtaining labeled data for training is laborious and time-consuming. Therefore, it is crucial to develop effective algorithms in situations with limited labeled samples. In this work, we propose a label-efficient deep learning method with self-ensembling (LESEN). LESEN incorporates supervised and unsupervised losses, enabling the student and teacher models to mutually learn from each other, forming a self-ensembling mean teacher framework. Additionally, we introduce a reliable unlabeled sample selection (RUSS) mechanism to further enhance LESEN's effectiveness. Our experiments on the human connectome project (HCP) dataset demonstrate the superior performance of our method when compared to state-of-the-art techniques, advancing multimodal VP segmentation for comprehensive analysis in clinical and research settings. The implementation code will be available at: https://github.com/aldiak/Semi-Supervised-Multimodal-Visual-Pathway- Delineation.
Paper Structure (9 sections, 10 equations, 4 figures, 1 table)

This paper contains 9 sections, 10 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Network Architecture
  4. Self-Ensembling Mean Teacher with RUSS
  5. Experiments and Results
  6. Dataset and Implementation details
  7. Results and Analysis
  8. Conclusion
  9. Acknowledgments

Figures (4)

  • Figure 1: The overall architecture of our proposed method. There are two subnetworks, including a student model and a teacher model with each model taking two inputs. The models share the same structure and are used to compute both supervised and unsupervised consistency losses during training.
  • Figure 2: The detailed structure of the LESEN's encoder.
  • Figure 3: Visualization of VP segmentation results produced by different methods.
  • Figure 4: Effectiveness of each component on the HCP dataset. Here w/o RUSS means without the RUSS method, w/o SA means without the spatial attention, and All represents our proposed method with all components.