Octascope: A Lightweight Pre-Trained Model for Optical Coherence Tomography.

Area of Science:

• Biomedical imaging
• Medical artificial intelligence
• Deep learning for medical image analysis

Background:

• Optical coherence tomography (OCT) provides high-resolution subsurface tissue imaging.
• Deep learning for OCT analysis faces challenges with limited training data and slow inference speeds.
• Developing efficient and accurate AI models for OCT is crucial for clinical applications.

Purpose of the Study:

• To develop a lightweight, domain-specific convolutional neural network (CNN) model for efficient and accurate OCT image analysis.
• To improve the generalizability of OCT analysis models across diverse tissue types.
• To achieve a balance between computational efficiency and diagnostic accuracy for real-time OCT applications.

Main Methods:

• Developed Octascope, a lightweight CNN model for OCT image analysis.
• Employed a curriculum learning approach for pre-training: natural images (ImageNet) followed by diverse OCT tissues (retinal, abdominal, renal).
• Evaluated Octascope on epidural tissue detection and retinal diagnosis tasks, comparing it against existing methods and a Transformer-based model.

Main Results:

• Octascope demonstrated improved accuracy in epidural tissue detection (9.13% over single-task learning, 5.95% over OCT-specific transfer learning).
• Octascope outperformed VGG16 (5.36%) and ResNet50 (6.66%) in retinal diagnosis.
• Octascope achieved 2 to 4.4 times faster inference speed than RETFound with comparable or better accuracy.

Conclusions:

• Octascope offers a significant advancement in OCT image analysis, balancing computational efficiency and diagnostic accuracy.
• The multi-domain pre-training strategy enhances model generalizability across different tissue types.
• Octascope is suitable for real-time clinical applications requiring rapid and reliable OCT image interpretation.