A Deep Learning Model for Cervical Optical Coherence Tomography Image Classification.

Area of Science:

• Gynecologic oncology
• Medical imaging
• Artificial intelligence in healthcare

Background:

• Optical coherence tomography (OCT) shows promise for in vivo cervical lesion detection, exceeding colposcopy's effectiveness.
• Gynecologists require advanced tools to interpret complex cervical OCT images due to unfamiliarity with the technology.
• Need for intelligent computer-aided diagnosis to improve efficiency and accuracy in cervical OCT image interpretation.

Purpose of the Study:

• To develop a clinically applicable deep learning (DL) model for classifying 3D OCT volumes of cervical tissue.
• To validate the DL model's efficacy in identifying high-risk cervical lesions, including high-grade squamous intraepithelial lesions and cervical cancer.

Main Methods:

• A convolutional neural network architecture incorporating a feature pyramid network (FPN) with texture encoding and deep supervision was developed.
• Extraction, representation, and fusion of four-scale texture features were performed to enhance classification of high-risk lesions.
• An auxiliary classification mechanism using deep supervision was implemented for adaptive FPN scale weighting and efficient model training.

Main Results:

• The DL model achieved an 81.55% F1-score, 82.35% sensitivity, and 81.48% specificity on the Renmin dataset, surpassing five medical experts.
• On the Huaxi dataset, the model attained an 84.34% F1-score, 87.50% sensitivity, and 90.59% specificity, comparable to top investigators.
• The DL model provides visual evidence of learned histomorphological and texture features, aiding gynecologists in rapid clinical decision-making.

Conclusions:

• The developed deep learning model demonstrates significant potential for efficient and effective cervical lesion screening using OCT.
• The AI tool can assist gynecologists by providing reliable interpretations of cervical OCT images, improving diagnostic capabilities.
• This approach offers a promising advancement in the early detection of cervical abnormalities and cancer.