Self-supervised learning with application for infant cerebellum segmentation and analysis

Yue Sun, Limei Wang, Kun Gao, Shihui Ying, Weili Lin, Kathryn L. Humphreys, Gang Li, Sijie Niu, Mingxia LiuNature Communications 2023 · DOI

TL;DR

This paper introduces a novel self-supervised learning (SSL) framework for accurately segmenting infant cerebellums, a crucial but challenging task due to low tissue contrast and data variability. The method significantly improves segmentation accuracy, enabling new insights into early cerebellar development, sex differences, and potential biomarkers for disorders like autism.

What Problem Does This Paper Address?

Accurately segmenting the infant cerebellum is vital for understanding early brain development, especially in the first two postnatal years, a period of rapid growth. However, this is hindered by several challenges:

  • Low Tissue Contrast: Especially in infants younger than three months, distinguishing between gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) is difficult, limiting the availability of manually annotated training data.
  • Dynamic Tissue Contrast: Tissue contrast changes significantly with age, further complicating segmentation models trained on specific age groups.
  • Inter-Site Data Heterogeneity (Domain Shift): Data acquired from different scanners, sites, and imaging parameters introduces variability that degrades the performance of models trained on one domain when applied to another.
  • Complex Anatomy: The cerebellum's tightly folded cortex and arbor vitae structure add to the segmentation complexity.

These challenges have limited our understanding of typical and disordered cerebellar development, particularly in the critical early postnatal period.

Methods

The authors propose a two-step Self-Supervised Learning (SSL) framework to overcome these limitations:

  1. Supervised Learning (Source Domain):
    • A segmentation model (ADU-Net) is initially trained on a dataset of 24-month-old infants. This age group was chosen because their MRIs exhibit better tissue contrast, allowing for more reliable manual annotations.
  2. A Confidence Model (ConM) is then trained. This model learns to predict the reliability of the automated segmentations generated by the ADU-Net at a voxel level. It uses tissue probability maps as input and is trained to distinguish between correctly and incorrectly segmented voxels based on their probability distributions. This ConM, trained on 24-month-old data, is then applied to younger infants.
  3. Self-Supervised Learning (Target Domain):
    • The key innovation here is using the Confidence Model to automatically generate reliable training samples for the target domain (e.g., younger infants or data from different sites). Voxels with high confidence scores are selected as reliable pseudo-labels.
  4. A spatially-weighted cross-entropy loss function is introduced. This loss function assigns higher weights to reliable voxels identified by the confidence map, effectively guiding the training of a new, domain-specific segmentation model for the target domain. This process mitigates the domain shift problem without requiring new manual annotations for each target domain.
  5. A gradual label propagation strategy is also employed, where models trained on older infants are used to generate pseudo-labels for progressively younger age groups, further refining the segmentation.

Datasets

The framework was validated on three datasets:

  • UNC/UMN Baby Connectome Project (BCP): This dataset provided the primary training and validation data, including labeled images from 24-month-olds (source domain) and unlabeled images across various ages (≤3, 6, 9, 12, 18 months) for testing the target domain adaptation.
  • Vanderbilt University: A small dataset of five 6-month-old infant MRIs acquired with a Philips scanner was used to test the framework's cross-site generalization capabilities.
  • National Database for Autism Research (NDAR): This dataset, containing longitudinal scans of 95 male subjects (22 autistic, 73 neurotypical) at 6, 12, and 24 months, was used to investigate developmental differences in autistic infants.

The paper also mentions validation on the iSeg-2019 challenge dataset, a multi-site infant cerebrum segmentation challenge.

Key Insights and Ideas

  • Leveraging Older Infants for Younger Ones: The core idea of using reliable segmentations from older infants (with better contrast) to train models for younger infants (with poor contrast) is a powerful transfer learning strategy.
  • Confidence Maps as Quality Control: The confidence model acts as an intelligent gatekeeper, identifying and prioritizing reliable segmentation predictions for training, which is crucial when ground truth is scarce.
  • Domain Adaptation via SSL: The SSL approach effectively bridges the "domain gap" between different imaging sites and age groups, leading to more robust and generalizable segmentation.
  • Early Development is Rapid and GM-Dominated: The first six months show the most dynamic cerebellar growth, with GM playing a more dominant role than WM.
  • Sex Differences in Cerebellar Growth: Cerebellar GM and WM volumes are consistently larger in males than females from birth to 24 months.
  • Autism and Cerebellar Growth: Autistic males appear to have larger cerebellar GM and WM volumes than neurotypical males, with potentially faster growth rates in early infancy.

Key Results

  • Improved Segmentation Accuracy: The proposed SSL method achieved superior segmentation accuracy (higher Dice ratios, lower HD95) compared to several state-of-the-art methods across different age groups and imaging sites. For example, for WM segmentation at ≤3 months, the Dice ratio for the proposed method was 90.39%, significantly higher than ADU-Net (83.96%) and other methods.
  • Robustness to Domain Shift: The method demonstrated strong generalization across different sites (Siemens vs. Philips scanners) and time points, outperforming other methods in cross-site evaluations.
  • Early Cerebellar Growth Dynamics: The study quantified the rapid growth in the first six months (GM: 141.58%, WM: 85.18%) and highlighted GM's dominant role.
  • Sex Differences: Statistically significant differences in cerebellar GM and WM volumes were observed between males and females, with large effect sizes.
  • Autism Biomarkers: Preliminary findings suggest potential differences in cerebellar volume and growth trajectories in autistic males compared to neurotypical males, particularly from 6 to 24 months.

Limitations and Future Work

  • Training Sample Size: The authors acknowledge that the training sample size, while improved, is still relatively small. They suggest using generative models to augment training data.
  • Cerebellum Atlas: The regular structure of the cerebellum could benefit from incorporating a cerebellum atlas as prior knowledge, especially for very young infants.
  • Longitudinal Subject Numbers: The conclusions regarding growth trajectories, particularly for autistic subjects, are based on a limited number of subjects and require further validation with larger cohorts.
  • Segmentation Granularity: The current work segments the cerebellum into GM, WM, and CSF. Future work aims to further parcellate the cerebellum into smaller regions of interest.
  • Modeling Appearance Trajectories: Incorporating varying tissue contrast at different time points could further boost performance.

Summary updated

Abstract

Accurate tissue segmentation is critical to characterize early cerebellar development in the first two postnatal years. However, challenges in tissue segmentation arising from tightly-folded cortex, low and dynamic tissue contrast, and large inter-site data heterogeneity have limited our understanding of early cerebellar development. In this paper, we propose an accurate self-supervised learning framework for infant cerebellum segmentation. We validate its accuracy using 358 subjects from three datasets. Our results suggest the first six months exhibit the most rapid and dynamic changes, with gray matter (GM) playing a dominant role in cerebellar growth over white matter (WM). We also find both GM and WM volumes are larger in males than females, and GM and WM volumes are larger in autistic males than neurotypical males. Application of our method to a larger population will fuel more cerebellar studies, ultimately advancing our comprehension of its structure and function in neurotypical and disordered development.
BibTeX
@inproceedings{sun2023self,
  author = {Yue Sun and Limei Wang and Kun Gao and Shihui Ying and Weili Lin and Kathryn L. Humphreys and Gang Li and Sijie Niu and Mingxia Liu and Li Wang},
  title = {{Self-supervised learning with application for infant cerebellum segmentation and analysis}},
  booktitle = {Nature Communications},
  year = {2023},
  doi = {10.1038/s41467-023-40446-z},
}
Figure