[Paper Review] Brain Tissue Segmentation Across the Human Lifespan via Supervised Contrastive Learning

Xiaoyang Chen, Jinjian Wu+7 more|arXiv (Cornell University)|Jan 3, 2023

Medical Image Segmentation Techniques4 citations

TL;DR

This paper proposes a unified deep learning framework for brain tissue segmentation across the entire human lifespan (0–100 years) using supervised contrastive learning to learn robust, discriminative feature representations. The method achieves state-of-the-art performance on a large dataset of 2,464 MR images, outperforming existing methods in segmenting white matter, gray matter, and cerebrospinal fluid despite wide anatomical and intensity variations across ages.

ABSTRACT

Automatic segmentation of brain MR images into white matter (WM), gray matter (GM), and cerebrospinal fluid (CSF) is critical for tissue volumetric analysis and cortical surface reconstruction. Due to dramatic structural and appearance changes associated with developmental and aging processes, existing brain tissue segmentation methods are only viable for specific age groups. Consequently, methods developed for one age group may fail for another. In this paper, we make the first attempt to segment brain tissues across the entire human lifespan (0-100 years of age) using a unified deep learning model. To overcome the challenges related to structural variability underpinned by biological processes, intensity inhomogeneity, motion artifacts, scanner-induced differences, and acquisition protocols, we propose to use contrastive learning to improve the quality of feature representations in a latent space for effective lifespan tissue segmentation. We compared our approach with commonly used segmentation methods on a large-scale dataset of 2,464 MR images. Experimental results show that our model accurately segments brain tissues across the lifespan and outperforms existing methods.

Motivation & Objective

• To develop a single deep learning model capable of accurately segmenting brain tissues across the entire human lifespan, from infancy to old age.
• To address the challenge of structural and contrast variability due to developmental and aging processes in MR images.
• To improve robustness to intensity inhomogeneity, motion artifacts, scanner differences, and diverse acquisition protocols.
• To overcome limitations of existing methods that are tailored for specific age groups (e.g., adults or infants) and fail when applied broadly.
• To demonstrate the effectiveness of supervised contrastive learning in learning discriminative features for lifelong segmentation.

Proposed method

• Uses a deep encoder-decoder network with a feature extractor and classifier head for voxel-wise segmentation.
• Applies supervised contrastive learning with a memory bank of prototype features per tissue class to enhance feature discrimination.
• Leverages cosine similarity between feature representations and class-specific proxies to pull same-class features together and push different classes apart.
• Combines cross-entropy loss with contrastive loss, weighted by a hyperparameter λ_ctr, to jointly optimize segmentation and feature learning.
• Employs a memory bank of size M to store representative features, reducing computational cost of pairwise comparisons.
• Uses temperature scaling τ to control the sharpness of the similarity distribution in the contrastive objective.

Experimental results

Research questions

• RQ1 Can a single deep learning model achieve accurate brain tissue segmentation across the entire human lifespan, from 0 to 100 years?
• RQ2 How effective is supervised contrastive learning in improving feature representation robustness to age-related anatomical and intensity changes in MR images?
• RQ3 To what extent does the model generalize across diverse acquisition protocols, scanners, and image quality conditions?
• RQ4 How does the inclusion of both T1- and T2-weighted images compare to single-modality input in segmentation performance?
• RQ5 What hyperparameter settings (M, τ, λ_ctr) yield optimal segmentation accuracy and stability?

Key findings

• The proposed method achieved a Dice Similarity Coefficient (DSC) of 94.2±1.1% for gray matter, 95.7±1.0% for white matter, and 88.5±3.8% for cerebrospinal fluid on the test set, outperforming existing methods.
• Balanced sampling of voxels across tissue classes improved DSC for minority classes, particularly CSF, increasing it from 83.0±6.2% (unbalanced) to 88.5±3.8% (balanced).
• Performance was only marginally reduced when using single T1- or T2-weighted images (DSC decrease of 0.4–1.9%), indicating strong generalization from limited input.
• Larger memory bank size (M=10) and smaller temperature (τ=0.1) yielded better segmentation results, consistent with contrastive learning theory.
• The model successfully captured nonlinear growth trajectories of tissue volumes: GM and WM increased then decreased, while CSF volume increased monotonically across the lifespan.
• High-quality white and pial surface reconstructions were generated, with clear cortical gyri and sulci, confirming detailed segmentation accuracy.