An explainable attention model for cervical precancer risk classification using colposcopic images

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Oncology

Background:

• Cervical cancer poses a significant global health challenge, with late-stage diagnosis leading to high mortality.
• Early detection and risk assessment are critical for effective cervical cancer prevention strategies.

Purpose of the Study:

• To develop and evaluate the Cervix-AID-Net model for classifying cervical precancer risk using colposcopic images.
• To enhance the interpretability of the AI model's decisions through integrated explainable AI techniques.

Main Methods:

• The Cervix-AID-Net model utilizes a Convolutional Block Attention Module (CBAM) and convolutional layers for feature extraction from DYSIS colposcopic images.
• Explainable AI methods including gradient class activation maps, Local Interpretable Model-agnostic Explanations, and CartoonX were integrated for decision transparency.
• The model was evaluated using holdout and ten-fold cross-validation techniques.

Main Results:

• The Cervix-AID-Net model achieved high classification accuracies of 99.33% (holdout) and 99.81% (ten-fold cross-validation).
• CartoonX provided detailed explanations by identifying relevant image regions, enhancing model interpretability.
• The model demonstrated robustness to moderate levels of Gaussian noise (up to 3%) and blur (up to 10%).

Conclusions:

• The Cervix-AID-Net model, incorporating CBAM and explainable AI, shows promise for improving cervical precancer risk assessment.
• This AI framework has the potential to significantly impact the early detection and prevention of cervical cancer, improving patient outcomes globally.