Weakly Supervised Deep Nuclei Segmentation With Sparsely Annotated Bounding Boxes for DNA Image Cytometry.

Area of Science:

• Biomedical Imaging
• Computational Pathology
• Computer Vision

Background:

• Nuclei segmentation is crucial for DNA ploidy analysis in cytopathology.
• Current fully supervised methods demand extensive and costly pixel-wise annotations.
• Image-based cytometry (DNA-ICM) offers objective DNA content measurement.

Purpose of the Study:

• To develop a novel weakly supervised nuclei segmentation framework.
• To reduce the reliance on pixel-wise labels in DNA ploidy analysis.
• To improve the efficiency and reduce the cost of nuclei segmentation for cytopathology.

Main Methods:

• Integration of traditional image segmentation and self-training for instance segmentation.
• Utilizing sparsely annotated bounding boxes instead of segmentation labels.
• Employing a teacher-student model architecture for mask refinement and pseudo-label generation.

Main Results:

• The proposed framework significantly outperforms existing weakly supervised methods.
• Effective nuclei segmentation achieved on both a custom DNA-ICM dataset and a public cytopathological dataset.
• Demonstrated the efficacy of using bounding boxes for nuclei segmentation in DNA ploidy analysis.

Conclusions:

• The novel weakly supervised approach offers a viable and efficient alternative to fully supervised nuclei segmentation.
• The method effectively leverages bounding box annotations for accurate DNA ploidy analysis.
• Public availability of code and dataset facilitates further research and application in cytopathology.