A deep learning based framework for accurate segmentation of cervical cytoplasm and nuclei

Area of Science:

• Medical Imaging
• Computational Biology
• Oncology

Background:

• Cervical cancer cell segmentation is crucial for diagnosis.
• Existing methods face challenges due to low contrast between background and cytoplasm.
• Accurate nucleus segmentation is key for reliable analysis.

Purpose of the Study:

• To propose a superpixel and convolution neural network (CNN) based method for cervical cancer cell segmentation.
• To address the challenge of low contrast in cytoplasm segmentation.
• To develop a coarse-to-fine nucleus segmentation strategy with refinement.

Main Methods:

• A superpixel and CNN-based approach for automated cell segmentation.
• Initial cytoplasm segmentation to overcome contrast issues.
• Deep learning for region of interest detection.
• A coarse-to-fine nucleus segmentation strategy with refinement.

Main Results:

• Achieved 94.50% accuracy for nucleus region detection.
• Obtained a precision of 0.9143±0.0202 for nucleus cell segmentation.
• Achieved a recall of 0.8726±0.0008 for nucleus cell segmentation.
• Demonstrated superior performance compared to related methods.

Conclusions:

• The proposed superpixel and CNN method effectively segments cervical cancer cells.
• The approach shows significant improvements in nucleus detection and segmentation accuracy.
• This method offers a promising tool for automated cervical cancer cell analysis.