Cervi-LLM for real time colposcopy lesion detection and interpretable diagnosis

Area of Science:

• Gynecologic Oncology
• Medical Imaging AI
• Digital Pathology

Background:

• Colposcopy is crucial for cervical cancer screening but requires specialized training, leading to variable diagnostic accuracy.
• Existing diagnostic performance in colposcopy varies significantly across different regions and skill levels.
• Need for an objective, accurate, and efficient tool to assist in colposcopic interpretation and diagnosis.

Purpose of the Study:

• To introduce Cervi-LLM, a multimodal Mixture of Experts (MoE) framework for assisting colposcopic detection of cervical lesions.
• To evaluate Cervi-LLM's performance in precise lesion localization and stratified diagnosis (Normal/LSIL/HSIL+).
• To compare Cervi-LLM's diagnostic capabilities against human physicians.

Main Methods:

• Developed Cervi-LLM, integrating YOLOMed for multi-modal image segmentation and a fine-tuned Large Language Model (LLM) for image-text analysis.
• Employed a two-level MoE architecture with dynamic gating, combining rule-based and data-driven approaches for diagnosis.
• Utilized a dataset comprising 126 cervical cancer, 692 HSIL, 306 LSIL, and 999 normal cases.

Main Results:

• Cervi-LLM achieved high performance in segmentation (PA 94.51%, meanIoU 78.23%) and classification (ACC 91.52%), outperforming U-Net and Polyp-PVT.
• Diagnostic accuracy for HSIL+ detection showed superior sensitivity (95.96%) and specificity (94.17%) compared to senior physicians (67.58%, 81.67%).
• The system provides real-time processing (~30 fps) and rapid diagnosis (0.30 ± 0.05 min).

Conclusions:

• Cervi-LLM demonstrates significant potential to overcome limitations in conventional colposcopy.
• The AI framework offers an accurate, real-time tool for cervical lesion screening and biopsy guidance.
• Cervi-LLM can enhance diagnostic consistency and improve patient outcomes in cervical cancer screening programs.