[Paper Review] Weakly Supervised Medical Image Segmentation

Pedro H. T. Gama, Hugo Oliveira+1 more|arXiv (Cornell University)|Aug 12, 2021

Radiomics and Machine Learning in Medical Imaging4 citations

TL;DR

This paper proposes a few-shot medical image segmentation method using sparse labels via a Model-Agnostic Meta-Learning (MAML)-based approach, enabling accurate dense segmentation predictions from less than 2% labeled pixels. The method achieves Jaccard scores comparable to fully supervised models, especially when target domains differ significantly from source domains.

ABSTRACT

In this paper, we propose a novel approach for few-shot semantic segmentation with sparse labeled images. We investigate the effectiveness of our method, which is based on the Model-Agnostic Meta-Learning (MAML) algorithm, in the medical scenario, where the use of sparse labeling and few-shot can alleviate the cost of producing new annotated datasets. Our method uses sparse labels in the meta-training and dense labels in the meta-test, thus making the model learn to predict dense labels from sparse ones. We conducted experiments with four Chest X-Ray datasets to evaluate two types of annotations (grid and points). The results show that our method is the most suitable when the target domain highly differs from source domains, achieving Jaccard scores comparable to dense labels, using less than 2% of the pixels of an image with labels in few-shot scenarios.

Motivation & Objective

• To reduce the annotation burden in medical image segmentation by leveraging sparse labeling in few-shot learning.
• To enable accurate dense segmentation predictions from minimal labeled data, particularly in low-resource medical imaging scenarios.
• To investigate the effectiveness of MAML-based meta-learning in adapting to domain shifts between source and target medical imaging datasets.
• To evaluate performance using two sparse annotation types: grid and point annotations.
• To achieve segmentation performance close to fully supervised models while using significantly fewer labeled pixels.

Proposed method

• The method employs Model-Agnostic Meta-Learning (MAML) to train a model that generalizes across few-shot episodes with sparse supervision.
• During meta-training, only sparse labels (grid or point annotations) are used to optimize the model's ability to predict dense segmentation masks.
• During meta-testing, the model is fine-tuned on support sets with sparse labels and tested on query sets requiring dense predictions.
• The model learns to infer full segmentation maps from limited, sparse annotations through meta-optimization of initialization parameters.
• The approach is evaluated on four Chest X-Ray datasets with varying domain characteristics to assess robustness under domain shift.
• The method uses Jaccard similarity as the primary evaluation metric for segmentation performance.

Experimental results

Research questions

• RQ1 Can a few-shot segmentation model trained on sparse labels achieve performance comparable to fully supervised models in medical imaging?
• RQ2 How does the proposed MAML-based method perform when the target domain differs significantly from the source domain?
• RQ3 What is the impact of annotation type (grid vs. points) on segmentation accuracy with minimal labeling?
• RQ4 To what extent can the model generalize to new, unseen medical imaging domains with only a few labeled examples?
• RQ5 How does the model perform when less than 2% of pixels are labeled, especially in low-data regimes?

Key findings

• The proposed method achieves Jaccard scores comparable to fully supervised models, even when using less than 2% of the pixels as labels.
• The method demonstrates strong generalization to target domains that differ significantly from source domains, outperforming baseline few-shot methods in such scenarios.
• Sparse grid annotations yield better performance than sparse point annotations, indicating that spatial distribution of labels affects model learning.
• The model maintains high segmentation accuracy across four diverse Chest X-Ray datasets, confirming its robustness to domain shift.
• The method significantly reduces annotation cost while maintaining high performance, making it suitable for real-world medical imaging applications.
• Meta-learning with sparse labels enables effective few-shot adaptation, demonstrating the feasibility of low-resource medical image segmentation.