Author: Mark Anastasio, Hua Li, Zhuchen Shao 👨🔬
Affiliation: Washington University School of Medicine, University of Illinois Urbana-Champaign 🌍
Purpose: Automated semantic segmentation of cell nuclei in microscopic images is vital for disease diagnosis and tissue microenvironment analysis. However, obtaining large annotated datasets for training is labor-intensive and prone to errors. Traditional semi-supervised methods often struggle with limited labeled data, as they rely on guidance from few annotations for model training. This results in inefficient use of unlabeled data and the inability of use diverse training datasets. We introduced DTSeg, a novel semi-supervised framework that integrates a latent diffusion model (LDM) with a transformer-based decoder to overcome these challenges for accurate cell nuclei segmentation.
Methods: In DTSeg framework, the LDM is pre-trained on unlabeled images only using forward/reverse diffusion processes, enabling it to learn robust features from the available data. With the pre-trained LDM and only a small amount of annotated data, a transformer decoder can be trained with limited labeled data only while aggregating multi-scale features extracted from the LDM for precise cell segmentation. By independently leveraging labeled and unlabeled data during training, DTSeg can effectively utilize non-annotated datasets from diverse nuclei types, avoiding task-specific constraints.
Results: Experiments on three diverse datasets demonstrated its superior performance across various scenarios with limited and diverse labeled data. DTSeg method also outperformed other state-of-the-art semi-supervised and supervised methods. Training the LDM independently on unlabeled datasets has proven to support DTSeg in achieving the highest mIoU and F1 scores, with statistically significant improvements (p < 0.05) compared to baseline methods. Visualization of cell segmentation, even for complex structures, further confirms its precision.
Conclusion: The DTSeg framework effectively tackles the challenge of limited and diverse labeled data by leveraging LDMs to extract multi-scale features from diverse unlabeled datasets. It demonstrates superior semi-supervised segmentation performance, particularly in scenarios with scarce labeled data, providing a robust solution for clinical and research applications in histopathology.