The Contrastive Network With Convolution and Self-Attention Mechanisms for Unsupervised Cell Segmentation.

Area of Science:

• Biomedical image analysis
• Computational pathology
• Deep learning applications

Background:

• Supervised cell instance segmentation requires time-consuming pixel-wise annotations.
• Weakly or semi-supervised methods reduce annotation burden but still need some labels.
• Fully unsupervised cell segmentation remains a challenge in biomedical imaging.

Purpose of the Study:

• To develop an end-to-end unsupervised deep learning model for cell instance segmentation.
• To segment individual cell regions from hematoxylin and eosin (H&E) stained slides without any annotation.
• To evaluate the model's performance and generalization ability compared to supervised methods.

Main Methods:

• An end-to-end unsupervised deep learning model was proposed.
• The model processes raw H&E stained slide data without requiring annotations or pseudo-label generation.
• Ablation experiments compared the proposed backbone with pure Convolutional Neural Networks (CNNs) and transformers.

Main Results:

• The unsupervised model demonstrated satisfactory performance in segmenting individual cell regions.
• The model exhibited strong generalization capabilities across various validation datasets.
• The proposed backbone showed superior performance in capturing object edge and context information compared to pure CNN or transformer architectures under the unsupervised approach.

Conclusions:

• The developed unsupervised model effectively segments cells in H&E images without annotations.
• The model offers a viable alternative to labor-intensive supervised methods in biomedical image analysis.
• The backbone architecture is effective for unsupervised cell segmentation, capturing crucial spatial and contextual information.