Segmentation of cervical cell images

Area of Science:

• Medical Imaging
• Computational Biology
• Cytopathology

Background:

• Automating cervical cytology screening presents challenges in accurate cell image segmentation.
• Effective segmentation is crucial for reliable classification of normal and abnormal cervical cells.

Purpose of the Study:

• To develop and evaluate a novel image segmentation technique for cervical cytology.
• To improve the accuracy of automated classification of cervical cell images.

Main Methods:

• A new segmentation thresholding method based on cell perimeter stability was developed.
• Structural information and contour analysis were used to separate cytoplasm from background.
• Simple clustering was employed to differentiate cytoplasm, folded cytoplasm, and nucleus.
• The Mahalanobis distance classifier was used for performance evaluation.

Main Results:

• Manual thresholding achieved 66.0% accuracy for 13 classes and 95.2% for normal-abnormal classification.
• The novel contour tracing technique yielded 52.9% and 90.0% accuracy, respectively.
• The developed methods were tested on 1500 cervical cell images.

Conclusions:

• The proposed segmentation method shows potential for enhancing automated cervical cytology screening.
• Accurate cell image segmentation is a critical step towards reliable automated diagnosis.