Attention-Enhanced Lightweight Architecture with Hybrid Loss for Colposcopic Image Segmentation.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Oncology

Background:

• Cervical cancer screening via computer-aided diagnosis faces segmentation challenges in colposcopic images.
• Inaccurate segmentation and incomplete boundary detection hinder diagnostic accuracy.

Purpose of the Study:

• To develop a lightweight segmentation model for enhanced accuracy and computational efficiency in colposcopic image analysis.
• To improve the precision of cervical cancer screening tools.

Main Methods:

• Integrated dual encoder backbones (ResNet50, MobileNetV2) for feature extraction.
• Utilized a lightweight atrous spatial pyramid pooling (ASPP) module for multi-scale context.
• Incorporated a novel attention module for enhanced feature detail and a refined decoder for boundary segmentation.

Main Results:

• Achieved high performance metrics: 97.56% training accuracy, 96.04% validation accuracy, 97.00% specificity, 96.78% sensitivity.
• Demonstrated superior segmentation with a 98.71% Dice coefficient and 97.56% IoU.
• Outperformed existing methods in colposcopic image segmentation accuracy and efficiency.

Conclusions:

• The proposed lightweight model offers precise and reliable colposcopic image segmentation.
• This advancement has potential for real-world clinical applications in cervical cancer diagnosis.
• The research contributes to improved diagnostic workflows and patient outcomes in oncology.