HyperGraph-based capsule temporal memory network for efficient and explainable diabetic retinopathy detection in...

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Diabetic retinopathy (DR) poses a significant risk of vision loss, necessitating accurate and early detection.
• Current deep learning models for DR detection face challenges with morphological variations, imaging inconsistencies, computational expense, and lack of interpretability.
• Existing models struggle with robustness to noisy data and real-world clinical deployment.

Purpose of the Study:

• To introduce the HyperGraph Capsule Temporal Network (HGCTN), a novel deep learning framework for accurate, scalable, and interpretable diabetic retinopathy detection.
• To address the limitations of current deep learning models in terms of computational cost, robustness, and interpretability.
• To develop a model capable of effectively tracking disease progression and facilitating real-world clinical application.

Main Methods:

• Developed HGCTN by integrating hypergraph neural networks for spatial relationships, capsule networks for hierarchical features, and a temporal capsular memory unit (TCMU) for temporal dependencies.
• Employed meta-learning techniques and noise injection strategies to enhance model adaptability and resilience to image variations.
• Validated HGCTN on the DRIVE and Diabetic Retinopathy datasets, comparing its performance against established models.

Main Results:

• HGCTN achieved superior accuracy, with the best results being 99.0% (HDCTN) and 98.8% (ADTATC), outperforming existing models like TAHDL (96.7%) and ADTATC (98.2%).
• Demonstrated exceptional recall (100% on DRIVE, 99.8% on Diabetic Retinopathy dataset) and high specificity (99.7% and 99.6% respectively), indicating minimal false negatives.
• Utilized hypergraph attention maps and capsule activation images to provide interpretable predictions for clinical audiences.

Conclusions:

• HGCTN establishes a new benchmark for diabetic retinopathy detection, offering high classification accuracy, reduced computational complexity, and improved generalization.
• The model's interpretability and robustness make it suitable for real-world clinical deployment in automated ophthalmic diagnosis systems.
• HGCTN effectively addresses key deficiencies in existing DR detection models, paving the way for advanced diagnostic tools.