Encoder-Weighted W-Net for Unsupervised Segmentation of Cervix Region in Colposcopy Images

Area of Science:

• Medical Imaging
• Computer-Aided Diagnostics
• Deep Learning

Background:

• Early diagnosis of cervical cancer significantly improves treatment outcomes.
• Colposcopy is a key screening tool for detecting cervical abnormalities, including cervical intraepithelial neoplasia (CIN).
• Accurate segmentation of the cervix region in colposcopic images is crucial for computer-aided diagnostic systems.

Purpose of the Study:

• To develop a fully unsupervised deep learning method for cervix region segmentation in colposcopy.
• To eliminate the need for image pre- and post-processing in cervix segmentation.
• To improve the performance of computer-aided diagnostics for cervical cancer.

Main Methods:

• A novel deep learning-based unsupervised method was proposed for cervix region identification.
• The method utilizes a modified W-Net architecture with a new loss function and scheduling scheme.
• No image pre- or post-processing steps were required.

Main Results:

• The proposed unsupervised method achieved a Dice coefficient of 0.71 for cervix segmentation.
• The method demonstrated superior performance with reduced computational cost compared to existing approaches.
• Segmentation masks generated by the method showed a significant reduction in outliers.

Conclusions:

• The unsupervised cervix segmentation method effectively identifies cervix regions in colposcopic images.
• This approach can enhance diagnostic performance for CIN detection and other cervical conditions.
• The study highlights the potential of unsupervised learning for improving colposcopy-based diagnostics.