CircWaveDL: Modeling of optical coherence tomography images based on a new supervised tensor-based dictionary...

Area of Science:

• Ophthalmic imaging analysis
• Medical image processing
• Machine learning for healthcare

Background:

• Optical Coherence Tomography (OCT) imaging is vital for detecting macular abnormalities.
• Traditional dictionary learning (DL) methods often ignore the multi-dimensional structure of OCT data.
• Tensor-based DL approaches offer a way to preserve this inherent data structure.

Purpose of the Study:

• To present a novel tensor-based DL algorithm, CircWaveDL, for enhanced OCT image classification.
• To leverage the multi-dimensional structure of OCT data using tensor modeling.
• To improve the accuracy and generalizability of OCT classification for macular diseases.

Main Methods:

• Developed CircWaveDL, a tensor-based DL algorithm modeling training data and dictionary as higher-order tensors.
• Utilized CircWave atoms for dictionary initialization and CANDECOMP/PARAFAC (CP) decomposition for tensor factorization.
• Learned class-specific sub-dictionaries and assigned test B-scans based on minimal residual error.

Main Results:

• CircWaveDL achieved high classification accuracies: 92.5% (Dataset 1), 86.1% (Dataset 2), and 89.3% (Dataset 3).
• The method demonstrated strong generalizability across three distinct external validation databases.
• Introduced a novel heatmap generation technique highlighting discriminative features for improved classification.

Conclusions:

• CircWaveDL effectively models the multi-dimensional structure of OCT data for superior classification performance.
• The tensor-based approach significantly outperforms previous OCT classification methods.
• The study validates CircWaveDL's efficacy and generalizability in clinical OCT image analysis.