Cluster-Re-Supervision: Bridging the Gap Between Image-Level and Pixel-Wise Labels for Weakly Supervised Medical...

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Weakly supervised learning reduces reliance on pixel-wise annotations for deep learning in medical image segmentation.
• Class activation maps (CAMs) are used for pixel-wise localization but often result in under- or over-segmentation due to image-level labels.
• Limited medical data exacerbates issues like under-segmentation (false negatives) and over-segmentation (false positives).

Purpose of the Study:

• To address under- and over-segmentation in weakly supervised medical image segmentation.
• To develop a novel paradigm for generating pixel-wise constraints from image-level labels.
• To refine CAMs for improved segmentation accuracy.

Main Methods:

• Proposed a "cluster-re-supervision" paradigm using unsupervised clustering to evaluate pixel contributions in CAMs.
• Generated pixel-wise supervision (clustering maps) to refine CAMs and reduce segmentation errors.
• Integrated self-supervised learning with an inter-modality image reconstruction module and random masking to enhance feature learning and stabilize clustering.

Main Results:

• Demonstrated superior performance of the proposed weakly-supervised framework on two public medical image segmentation datasets.
• Successfully reduced both under-segmentation (false negatives) and over-segmentation (false positives).
• Validated the effectiveness of cluster-re-supervision in refining CAMs.

Conclusions:

• The proposed cluster-re-supervision framework significantly improves weakly supervised medical image segmentation.
• The method effectively generates pixel-wise constraints from image-level labels, overcoming limitations of traditional CAM-based approaches.
• Cluster-re-supervision is task-independent and extensible to other applications.