FBA-Net: Foreground and Background Aware Contrastive Learning for Semi-Supervised Atrium Segmentation

Yunsung Chung, Chanho Lim, Chao Huang, Nassir Marrouche, Jihun Hamm

TL;DR

This paper introduces FBA-Net, a novel semi-supervised learning approach for medical image segmentation that leverages contrastive learning to explicitly learn distinct representations for foreground and background regions. By focusing on these representations, FBA-Net achieves state-of-the-art performance with significantly reduced labeled data requirements, making medical image analysis more efficient and accessible.

What Problem Does This Paper Address?

Accurate segmentation of anatomical structures in medical images, such as the left atrium (LA) in cardiac MRI, is crucial for diagnosis, treatment planning, and disease monitoring. However, creating these segmentations manually is a laborious and time-consuming process that demands specialized expertise. While deep learning methods have shown great promise, they typically require vast amounts of labeled data, which exacerbates the annotation bottleneck. Semi-supervised learning, which utilizes both labeled and unlabeled data, offers a solution, and contrastive learning has emerged as a powerful technique within this paradigm. However, existing contrastive learning methods for medical image segmentation often focus on voxel-wise relationships or rely heavily on data augmentation, which can be computationally expensive and may not fully exploit the specific characteristics of segmentation tasks. This paper aims to bridge this gap by proposing a more efficient and effective contrastive learning strategy tailored for medical image segmentation.

Methods

The core innovation of FBA-Net lies in its Foreground and Background Aware (FBA) contrastive learning strategy. Instead of relying solely on data augmentation to create positive pairs, FBA-Net explicitly extracts representations for both the foreground (target object) and background regions within an image. This is achieved by using an encoder that maps these distinct regions into a latent space.

The contrastive learning module then trains the network to:

  • Pull foreground representations closer to other foreground representations (and similarly for background).
  • Push foreground and background representations further apart.

This explicit focus on differentiating foreground and background helps the model learn more robust and discriminative features, leading to better boundary delineation. The paper also introduces a novel ranking-based weighting scheme for the contrastive loss, which allows for effective training with smaller batch sizes, thus reducing computational costs compared to methods that require large batches for optimal contrastive learning performance. The overall loss function combines this contrastive loss with a standard Dice loss for labeled data and a consistency loss for unlabeled data.

Datasets

The proposed FBA-Net was evaluated on three publicly available medical image segmentation datasets:

  • LA Dataset: A benchmark dataset from the 2018 MICCAI Atria Segmentation Challenge, consisting of 3D GE CMR images for left atrium segmentation.
  • Pancreas-CT Dataset: A dataset of 3D abdominal contrast-enhanced CT scans.
  • ACDC Dataset: A collection of cardiac MRIs containing left and right ventricles, and myocardium.

Key Insights and Ideas

  • Explicit Foreground-Background Contrast is Key: The paper demonstrates that directly learning to distinguish between foreground and background representations, rather than relying solely on general data augmentation, significantly improves segmentation performance, especially in semi-supervised settings.
  • Efficiency through Targeted Contrast: By focusing contrastive learning on foreground and background regions, FBA-Net reduces the reliance on computationally intensive data augmentation techniques often used in other contrastive learning methods.
  • Small Batch Size Effectiveness: The introduction of ranking weights for the contrastive loss allows the model to achieve high performance even with small batch sizes, making the method more accessible and less computationally demanding.
  • Versatile Applicability: The method shows strong performance across different datasets and segmentation tasks (binary and multi-class), suggesting its broad utility in medical image analysis.

Key Results

FBA-Net achieves state-of-the-art performance on the evaluated datasets, particularly in semi-supervised settings.

  • On the LA dataset, with only 20% labeled data, FBA-Net achieved a Dice score of 91.31%, which is remarkably close to the fully supervised performance of 91.62% (using 100% labeled data).
  • The method consistently outperforms other state-of-the-art semi-supervised methods, including CVRL and RCPS, across various metrics (Dice and ASD) on the LA, Pancreas-CT, and ACDC datasets.
  • Qualitative results show that FBA-Net produces highly accurate segmentations that closely match ground truth, even in challenging areas, outperforming other methods in terms of both over-segmentation and under-segmentation.
  • An ablation study showed that incorporating the FBA contrastive module into other semi-supervised methods led to significant improvements in Dice scores.

Limitations and Future Work

  • Limitations: The paper mentions that contrastive learning with large batch sizes can lead to faster learning, but their proposed method aims to achieve effectiveness with smaller batches. While this is an advantage, it implies that larger batch sizes might still offer further improvements, albeit at a higher computational cost.
  • Future Work: The authors suggest that the proposed approach can be extended to other medical image segmentation tasks where foreground-background separation is crucial for accurate segmentation. This implies potential applications beyond cardiac imaging and left atrium segmentation.

Summary updated

Abstract

Medical image segmentation of gadolinium enhancement magnetic resonance imaging (GE MRI) is an important task in clinical applications. However, manual annotation is time-consuming and requires specialized expertise. Semi-supervised segmentation methods that leverage both labeled and unlabeled data have shown promise, with contrastive learning emerging as a particularly effective approach. In this paper, we propose a contrastive learning strategy of foreground and background representations for semi-supervised 3D medical image segmentation (FBA-Net). Specifically, we leverage the contrastive loss to learn representations of both the foreground and background regions in the images. By training the network to distinguish between foreground-background pairs, we aim to learn a representation that can effectively capture the anatomical structures of interest. Experiments on three medical segmentation datasets demonstrate state-of-the-art performance. Notably, our method achieves a Dice score of 91.31% with only 20% labeled data, which is remarkably close to the 91.62% score of the fully supervised method that uses 100% labeled data on the left atrium dataset. Our framework has the potential to advance the field of semi-supervised 3D medical image segmentation and enable more efficient and accurate analysis of medical images with a limited amount of annotated labels.
BibTeX
@inproceedings{chung2023fba,
  author = {Yunsung Chung and Chanho Lim and Chao Huang and Nassir Marrouche and Jihun Hamm},
  title = {{FBA-Net: Foreground and Background Aware Contrastive Learning for Semi-Supervised Atrium Segmentation}},
  booktitle = {MILLanD @ MICCAI},
  year = {2023},
  url = {https://openreview.net/forum?id=CT3OpNTbHX},
}
Figure