Dual supervised sampling networks for real-time segmentation of cervical cell nucleus

Area of Science:

• Computational pathology
• Medical image analysis
• Deep learning for cancer diagnostics

Background:

• Accurate segmentation of cervical cell nuclei is critical for pathological identification and classification.
• Existing deep learning methods, while accurate, are computationally intensive for whole-slide images (WSIs).

Purpose of the Study:

• To develop an efficient deep learning model for precise cervical cell nucleus segmentation.
• To significantly reduce the computational cost and inference time for analyzing large-scale cytological datasets.

Main Methods:

• Proposed a dual-supervised sampling network (DSSNet) incorporating a supervised-down sampling module using compressed images.
• Integrated a boundary detection network to guide the up-sampling process for enhanced segmentation accuracy.
• Implemented strategies to reduce convolutional calculations during feature extraction.

Main Results:

• DSSNet achieved a 5-fold increase in inference speed compared to UNet on various cervical cell datasets.
• The proposed method maintained segmentation accuracy comparable to existing state-of-the-art approaches.
• Demonstrated the effectiveness of dual-supervision and compressed image processing for efficient nucleus segmentation.

Conclusions:

• The dual-supervised sampling network offers a significant speed improvement for cervical cell nucleus segmentation.
• This approach addresses the computational challenges of analyzing large whole-slide images in digital pathology.
• The developed method provides a viable solution for faster and accurate pathological cell identification.