This paper presents a comprehensive survey on the recent progress in weakly supervised medical image segmentation. It covers various forms of weak annotations, including image-level labels, bounding boxes, scribbles, and points, and discusses how these annotations can be leveraged to train deep learning models for segmentation tasks.
The key highlights and insights from the survey are:
Image-level annotations: Techniques like class activation mapping and iterative pseudo-mask generation can be used to bridge the supervision gap between image-level labels and pixel-wise segmentation.
Bounding box annotations: Methods that incorporate bounding box tightness priors and smooth maximum approximation can effectively utilize bounding box information for segmentation.
Scribble annotations: Approaches that leverage scribble information to guide the segmentation process, either through loss function constraints or pseudo-label generation, have shown promising results.
Point annotations: Techniques that enforce inequality constraints with differentiable penalties or leverage contextual regularization can effectively utilize sparse point annotations for segmentation.
Partially supervised datasets: Strategies that combine fully and partially supervised learning, such as pseudo-label generation and multi-task learning, can address the challenge of limited labeled data.
The emergence of foundation models, particularly the Segment Anything Model (SAM), has introduced innovative capabilities for segmentation tasks using weak annotations, enabling more efficient and scalable medical image segmentation.
The survey also discusses several challenges and potential solutions, such as quality evaluation and control of weak annotations, integrating domain knowledge, and leveraging existing datasets, to further advance the field of weakly supervised medical image segmentation.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Yuyan Shi,Ji... at arxiv.org 04-23-2024
https://arxiv.org/pdf/2404.13239.pdfDeeper Inquiries