Cross-Attention Based Multi-Resolution Feature Fusion Model for Self-Supervised Cervical OCT Image Classification.

Area of Science:

• Biomedical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Cervical cancer poses a significant threat to women's reproductive health and lives.
• Optical Coherence Tomography (OCT) offers non-invasive, high-resolution imaging of cervical tissues.
• Supervised learning for cervical OCT image analysis is challenged by the difficulty in obtaining large, labeled datasets.

Purpose of the Study:

• To develop a computer-aided diagnosis (CADx) approach for effective cervical OCT image classification.
• To leverage a self-supervised Vision Transformer (ViT) model for improved transfer learning capabilities.
• To enhance the detection and localization of cervical lesions through interpretable AI.

Main Methods:

• Utilized Masked Autoencoders (MAE) for self-supervised pre-training of the ViT model on cervical OCT images.
• Employed a ViT-based classification model with a cross-attention module for multi-scale feature fusion.
• Conducted ten-fold cross-validation on a multi-center dataset and external validation on 3D OCT volumes.

Main Results:

• Achieved an AUC of 0.9963 ± 0.0069, with 95.89 ± 3.30% sensitivity and 98.23 ± 1.36% specificity in binary classification.
• Demonstrated superior performance compared to state-of-the-art Transformer and CNN models.
• External validation yielded 92.06% sensitivity and 95.56% specificity, meeting or exceeding expert performance.

Conclusions:

• The self-supervised ViT-based CADx model effectively classifies cervical OCT images for high-risk disease detection.
• The model exhibits strong transfer learning ability and interpretability via attention maps.
• This approach shows significant potential for improving gynecological diagnostics and patient outcomes.