A new convolutional neural network based on combination of circlets and wavelets for macular OCT classification

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Artificial intelligence (AI) algorithms, including machine learning and deep learning, show promise in analyzing retinal optical coherence tomography (OCT) images for early ocular abnormality detection.
• A significant limitation in medical AI is the scarcity of data, which can hinder diagnostic accuracy.
• Time-frequency transforms offer a potential solution to enhance AI performance in medical image analysis.

Purpose of the Study:

• To investigate the impact of non-data-adaptive time-frequency transforms (X-lets) on the classification of OCT B-scans.
• To develop a novel transform that improves the simultaneous classification accuracy of both normal and abnormal retinal OCT images.
• To enhance the interpretability of AI models in ophthalmology.

Main Methods:

• Employed various X-lets to transform OCT B-scans individually, using all resulting sub-bands as input for a 2D Convolutional Neural Network (CNN).
• Designed a novel CircWave transform by concatenating sub-bands from the 2D Discrete Wavelet Transform (2D-DWT) and the circlet transform.
• Evaluated classification performance on two independent datasets from different imaging systems.

Main Results:

• The 2D-DWT excelled in classifying normal cases, while the circlet transform was superior for abnormal cases with circular fluid accumulation.
• The proposed CircWave transform significantly outperformed individual transforms and original images in per-class accuracy for both normal and abnormal cases.
• Grad-CAM visualization demonstrated that CircWaveNet focused on relevant circular formations in abnormal cases and linear structures in normal cases, unlike original B-scans.
• Achieved high accuracies of 94.5% and 90% on two distinct datasets, demonstrating generalizability.

Conclusions:

• The CircWave transform, integrated into the CircWaveNet CNN architecture, offers superior diagnostic accuracy for ocular abnormalities detected via OCT.
• This approach enhances the interpretability of AI models by focusing on diagnostically relevant features.
• The CircWave transform represents a significant advancement in AI-assisted ophthalmic diagnostics, addressing data limitations and improving classification performance.