DeepRGVP: A Novel Microstructure-Informed Supervised Contrastive Learning Framework for Automated Identification Of The Retinogeniculate Pathway Using dMRI Tractography

Sipei Li, Jianzhong He, Tengfei Xue, Guoqiang Xie, Shun Yao, Yuqian Chen, Erickson F. Torio, Yuanjing Feng, Dhiego Ca Bastos, Yogesh Rathi, Nikos Makris, Ron Kikinis, Wenya Linda Bi, Alexandra J Golby, Lauren J O'Donnell, Fan ZhangarXiv 2022 · arXiv

TL;DR

DeepRGVP is the first deep learning framework designed to automatically identify the retinogeniculate pathway (RGVP)—the critical visual link between the retina and the brain—from dMRI tractography. By integrating tissue microstructure data into a supervised contrastive learning objective and employing a novel streamline-level data augmentation strategy, the model achieves state-of-the-art accuracy in isolating this complex anatomical structure.

What Problem Does This Paper Address?

The retinogeniculate pathway (RGVP) is the "highway" for visual information, consisting of the optic nerve, optic chiasm, and optic tract. Accurately identifying this pathway is vital for neurosurgery and studying visual diseases. Currently, researchers rely on diffusion MRI (dMRI) tractography, which reconstructs 3D "streamlines" representing nerve fibers.

However, identifying the RGVP within these massive datasets is a bottleneck. It currently requires experts to manually draw regions of interest (ROIs) to filter streamlines—a process that is slow, expensive, and prone to human error. Automated deep learning methods exist for the brain's cerebrum, but the RGVP presents two unique challenges:

  1. Geometric Ambiguity: RGVP streamlines look nearly identical in shape and position to neighboring non-RGVP streamlines, making spatial coordinates alone insufficient for classification.
  2. Class Imbalance: In a typical dataset, RGVP streamlines are outnumbered by others at a ratio of roughly 1:8, making it difficult for standard models to learn their specific features.

Methods

DeepRGVP builds upon a point-cloud-based network architecture (SupWMA) and introduces two major innovations:

  • MicroSCL (Microstructure-informed Supervised Contrastive Learning): Traditional contrastive learning groups data points based solely on class labels. MicroSCL goes a step further by incorporating Fractional Anisotropy (FA)—a measure of tissue microstructure. The model is forced to learn that "positive" pairs of streamlines must not only share a label but also have similar FA values. This helps the network distinguish between streamlines that look the same geometrically but have different biological properties.
  • StreamDA (Streamline-level Data Augmentation): To solve the imbalance problem, the authors developed a subsampling-based augmentation. Instead of just duplicating data, they represent each streamline with a set of points and then randomly subsample different subsets of those points to create "new" training examples. This provides the model with more varied representations of the minority class (RGVP).
Fig. 1. Overview of DeepRGVP. (a) Example input tractography data and ROIs used for generating ground truth data. (b) Graphic illustration of the StreamDA method. (c) The overall network architecture, including the MicroSCL network and the downstream classifier for streamline classification.

Datasets

The study utilized high-quality dMRI data from the Human Connectome Project (HCP).

  • Initial Pool: 100 subjects.
  • Final Dataset: 62 subjects (38 were excluded due to incomplete coverage of the RGVP caused by face-removal anonymization).
  • Split: 40 subjects for training, 10 for validation, and 12 for testing.
  • Ground Truth: Expert-drawn and multi-rater validated ROIs were used to label streamlines as RGVP or non-RGVP.

Key Insights and Ideas

  • Biology as a Feature: The most significant "aha moment" is that geometry isn't enough in the crowded environment of the skull base. By treating diffusion microstructure (FA) as a fundamental feature rather than just an output, the model can disambiguate bundles that are physically touching but anatomically distinct.
  • Smart Subsampling: Rather than adding synthetic noise (which can distort anatomical truth), StreamDA uses the inherent redundancy in streamline points. This effectively "jitters" the input in a way that respects the underlying anatomy while providing the necessary volume of data for deep learning.
  • Contrastive Clustering: By using a contrastive loss, the model learns a global feature space where RGVP streamlines naturally cluster together, making the final classification task much simpler for the downstream layers.

Key Results

DeepRGVP outperformed several state-of-the-art (SOTA) methods, including DeepWMA, DCNN, and SupWMA.

  • Quantitative Performance: As shown in the table below, DeepRGVP achieved the highest Accuracy (96.646%) and F1 Score (84.316%) when using the augmented dataset. Quantitative comparisons with SOTA
Quantitative comparisons with SOTA
  • Ablation Success: The ablation study confirmed that adding microstructure information (MicroSCL) and the subsampling augmentation (StreamDA) both provided incremental boosts to precision and recall. Ablation study results
Ablation study results
  • Qualitative Superiority: Visually, DeepRGVP was less "overinclusive" than other models. It successfully identified fine structures, such as the ipsilateral pathway through the optic chiasm, which other models often missed or blurred. Visual comparison of identified RGVPs
Fig. 2. Visual comparison of the identified RGVPs in two example subjects. The inset images are provided for better visualization of local RGVP regions.

Limitations and Future Work

While the results are promising, the authors note a few areas for further development:

  • Pathological Data: The model was trained and tested on healthy subjects from the HCP. Its performance on patients with brain tumors or lesions that physically displace or degrade the RGVP remains to be seen.
  • Generalization: Future work will involve testing the framework on different populations and potentially extending the "microstructure-informed" approach to identify other cranial nerves that are similarly difficult to isolate.

Summary updated

Abstract

The retinogeniculate pathway (RGVP) is responsible for carrying visual information from the retina to the lateral geniculate nucleus. Identification and visualization of the RGVP are important in studying the anatomy of the visual system and can inform treatment of related brain diseases. Diffusion MRI (dMRI) tractography is an advanced imaging method that uniquely enables in vivo mapping of the 3D trajectory of the RGVP. Currently, identification of the RGVP from tractography data relies on expert (manual) selection of tractography streamlines, which is time-consuming, has high clinical and expert labor costs, and affected by inter-observer variability. In this paper, we present what we believe is the first deep learning framework, namely DeepRGVP, to enable fast and accurate identification of the RGVP from dMRI tractography data. We design a novel microstructure-informed supervised contrastive learning method that leverages both streamline label and tissue microstructure information to determine positive and negative pairs. We propose a simple and successful streamline-level data augmentation method to address highly imbalanced training data, where the number of RGVP streamlines is much lower than that of non-RGVP streamlines. We perform comparisons with several state-of-the-art deep learning methods that were designed for tractography parcellation, and we show superior RGVP identification results using DeepRGVP.
BibTeX
@misc{li2022deeprgvp,
  author = {Sipei Li and Jianzhong He and Tengfei Xue and Guoqiang Xie and Shun Yao and Yuqian Chen and Erickson F. Torio and Yuanjing Feng and Dhiego CA Bastos and Yogesh Rathi and Nikos Makris and Ron Kikinis and Wenya Linda Bi and Alexandra J Golby and Lauren J O'Donnell and Fan Zhang},
  title = {{DeepRGVP: A Novel Microstructure-Informed Supervised Contrastive Learning Framework for Automated Identification Of The Retinogeniculate Pathway Using dMRI Tractography}},
  year = {2022},
  eprint = {2211.08119},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2211.08119},
}
Figure