Semi-supervised semantic segmentation of cell nuclei with diffusion model and collaborative learning.

Area of Science:

• Biomedical image analysis
• Computational pathology
• Machine learning for medical imaging

Background:

• Automated cell nuclei segmentation is vital for disease diagnosis and tissue analysis.
• Acquiring large labeled datasets for supervised learning is challenging.
• Semi-supervised methods offer a solution by leveraging unlabeled data.

Purpose of the Study:

• To develop an effective semi-supervised learning framework for cell nuclei segmentation.
• To address challenges posed by extremely limited labeled data or diverse annotation types.
• To improve segmentation performance on both in-distribution and out-of-distribution (OOD) data.

Main Methods:

• Introduced a semi-supervised framework combining a latent diffusion model (LDM) with a transformer-based decoder.
• Utilized unsupervised LDM training on diverse unlabeled datasets for feature extraction.
• Employed a sequential training strategy and explored a collaborative learning approach.
• The framework, DTSeg, supports multi-channel inputs.

Main Results:

• The proposed DTSeg framework significantly outperformed existing semi-supervised and supervised methods on four diverse datasets.
• Achieved consistent performance across varying cell types and different amounts of labeled data.
• Demonstrated strong generalization capabilities for both in-distribution and OOD scenarios.

Conclusions:

• The DTSeg framework effectively performs cell nuclei segmentation with limited labeled data by integrating unsupervised LDM training.
• Collaborative learning enhances generalization, leading to superior results across diverse datasets.
• The method shows strong adaptability and generalization to various data conditions.