Abstract
Delineating the human visual pathway (VP) is a high-stakes task for neurosurgery and diagnosis, yet it is hindered by extreme anatomical variability and a lack of expert-labeled data. This paper introduces a semi-supervised framework utilizing Correlation-constrained Feature Decomposition (CFD) to intelligently fuse multi-parametric MRI sequences and a Consistency-based Sample Enhancement (CSE) module to optimize learning from unlabeled data. The resulting model significantly outperforms state-of-the-art methods, maintaining robust performance even when trained on a single labeled subject.
Introduction
The visual pathway—comprising the optic nerve, chiasm, and optic tract—serves as the primary conduit for visual data from the eyes to the brain. Accurate mapping of these structures is vital for avoiding surgical complications and treating visual disorders. However, two major hurdles persist:
- Anatomical Complexity: The VP varies significantly in shape and orientation between individuals. While multi-parametric MRI (e.g., T1-weighted structural scans and Fractional Anisotropy (FA)) provides complementary data, standard models often struggle to integrate these sequences effectively.
- Labeling Bottleneck: Expert manual delineation is labor-intensive, leading to a surplus of raw MRI data but a deficit of ground-truth labels.
Methodology
Correlation-constrained Feature Decomposition (CFD)
Rather than simply concatenating T1 and FA images, the CFD module decomposes each sequence into unique and non-unique features. Non-unique features are shared across sequences and facilitate alignment, while unique features capture sequence-specific characteristics (e.g., structural edges in T1 vs. fiber connectivity in FA). By prioritizing unique features in the final segmentation stage, the model avoids "feature entanglement" that typically blurs boundaries in multi-parametric imaging.
Consistency-based Sample Enhancement (CSE)
To leverage unlabeled data, the authors adopt a Mean Teacher architecture. A "Student" model learns from labeled data, while a "Teacher" model (an exponential moving average of the student) generates pseudo-labels for unlabeled samples. The CSE module acts as a quality controller by calculating an "inconsistency score" between the two models. Samples with high disagreement are discarded, ensuring the model learns only from high-confidence samples and preventing the degradation of the training process by noise.
Key Insights and Results
Feature Separation
The efficacy of the CFD module is demonstrated through t-SNE visualizations. The analysis reveals that "unique" features create a significantly sharper separation between the visual pathway and the background compared to non-unique or combined features.
Performance Benchmarks
The framework was validated on the Human Connectome Project (HCP) dataset and an in-house Multi-Shell Diffusion MRI (MDM) dataset:
- Superior Accuracy: On the HCP dataset, the proposed method achieved a Dice Score (DSC) of 0.85, outperforming the baseline (0.75) and other semi-supervised SOTA methods (0.77–0.79).
- Extreme Low-Data Robustness: When trained on the MDM dataset with only one labeled subject, the model achieved a DSC of 0.70, whereas the baseline model collapsed to 0.34.
- Structural Integrity: Qualitative analysis indicates that the framework maintains the continuity of thin structures like the optic tract, which are often fragmented by other models.
Conclusion and Future Work
This research demonstrates that intelligent feature decomposition and selective pseudo-labeling can overcome the data scarcity challenges in medical image segmentation. While currently a 2D-based architecture, future iterations may transition to native 3D frameworks to capture more spatial context. Additionally, moving from delineation to clinical quantification will be essential for translating these findings into routine diagnostic practice.
Summary updated