Classification of cervical neoplasms on colposcopic photography using deep learning

Area of Science:

• Gynecology
• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Colposcopy is crucial for cervical cancer detection but requires specialized expertise, which is limited in developing regions.
• Artificial intelligence (AI) offers potential solutions for computer-aided diagnosis, enhancing accuracy and accessibility.
• Automated analysis of colposcopic images could address the shortage of skilled physicians.

Purpose of the Study:

• To develop and validate deep learning models for the automated classification of cervical neoplasms using colposcopic photographs.
• To evaluate model performance using two established grading systems: cervical intraepithelial neoplasia (CIN) and lower anogenital squamous terminology (LAST).
• To assess the models' ability to discriminate between high-risk and low-risk lesions and identify lesions requiring biopsy.

Main Methods:

• Fine-tuning pre-trained convolutional neural networks (Inception-Resnet-v2 and Resnet-152) on a dataset of colposcopic images.
• Classifying lesions according to the CIN and LAST systems.
• Calculating multi-class classification accuracies and Area Under the Curve (AUC) for diagnostic performance.
• Utilizing attention maps to visualize model predictions.

Main Results:

• Resnet-152 achieved accuracies of 51.7% for the CIN system and 74.7% for the LAST system.
• AUC for discriminating high-risk from low-risk lesions was 0.781 (CIN) and 0.708 (LAST) using Resnet-152.
• Resnet-152 demonstrated high efficiency in detecting lesions requiring biopsy, with an AUC of 0.947.
• Attention maps provided meaningful insights into the model's decision-making process.

Conclusions:

• Deep learning models show potential for automated analysis of colposcopic images in cervical cancer screening.
• AI-powered tools could augment the diagnostic capabilities of healthcare providers, particularly in resource-limited settings.
• Further validation and integration of AI could improve the accuracy and efficiency of cervical neoplasm detection.