A multimodal deep learning model for cervical pre-cancers and cancers prediction: Development and internal validation...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Current cervical cancer screening methods lack objectivity and reproducibility.
• Limitations in existing diagnostic approaches necessitate improved accuracy and consistency.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) based risk prediction model for cervical neoplasia.
• To assess the clinical impact of the DL model in reducing unnecessary procedures.

Main Methods:

• A DL model was developed and validated using retrospective data from 6356 patients undergoing LEEP-conization/cone-biopsy.
• The model integrated clinical data and colposcopy images for predicting CIN2+ status.
• Performance was compared against expert clinician impressions.

Main Results:

• The DL model combining clinical history and images achieved high accuracy (90.8%) and AUC-ROC (95.3%) in predicting CIN2+.
• The model outperformed predictions based on image or clinical data alone, and clinician impressions.
• Applying a 10% probability threshold could avoid up to 35% of conizations without missing CIN2+ cases.

Conclusions:

• A novel DL model demonstrates significant potential for predicting cervical neoplasia and reducing unnecessary conizations.
• Further external validation is recommended to confirm the generalizability of the DL model.