Classification of cervical lesions based on multimodal features fusion

Area of Science:

• Gynecologic Oncology
• Medical Imaging Analysis
• Machine Learning for Healthcare

Background:

• Cervical cancer poses a significant global health threat, necessitating effective early detection methods.
• Early screening is crucial for preventing and treating cervical cancer, given its long development cycle and known causes.

Purpose of the Study:

• To develop a robust four-category classification model for cervical lesions (Normal, LSIL, HSIL, Ca).
• To integrate diverse data modalities, including colposcopy images, segmentation masks, HPV, TCT, and age, for enhanced diagnostic accuracy.

Main Methods:

• Utilized original and acetic colposcopy images, incorporating acetowhite opacity analysis to correlate image features with lesion grades.
• Employed lesion segmentation masks to integrate prior knowledge of lesion location and shape.
• Developed a cross-modal feature fusion module with self-attention to combine image and clinical text data.

Main Results:

• The proposed model demonstrated superior classification performance compared to existing models.
• Ablation studies confirmed that each component of the model independently enhances classification accuracy.
• The model achieved a favorable balance between performance and computational complexity.

Conclusions:

• The developed model shows significant promise for improving the accuracy and efficiency of cervical cancer screening.
• Integrating multi-modal data and advanced fusion techniques can enhance the diagnostic capabilities for cervical lesions.
• Further validation of this model could lead to improved clinical decision-making in gynecologic oncology.