Domain Anchored Features for Classification of OCT Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) is vital for diagnosing retinal diseases.
• Accurate OCT image classification aids in personalized treatment strategies.
• Existing deep learning models face challenges like feature congestion and misclassification of normal retinas.

Purpose of the Study:

• To develop an innovative deep neural network module for enhanced OCT image classification.
• To address feature congestion and improve the distinction between retinal diseases and normal retinas.
• To improve the accuracy of classifying eight types of retinal diseases and normal retinas.

Main Methods:

• Proposed a novel deep neural network module to enhance imaging features for distinct classification.
• Introduced a cross-domain feature anchoring strategy based on medical findings.
• Evaluated the model on two datasets for eight-class and four-class classification tasks.

Main Results:

• The proposed module significantly outperformed state-of-the-art methods in OCT image classification.
• Experimental results validated the effectiveness of the enhanced features for improved diagnostic accuracy.
• Ablation studies and sensitivity tests confirmed the robustness and comprehensive evaluation of the method.

Conclusions:

• The novel deep neural network module effectively enhances OCT image features for superior classification accuracy.
• The proposed approach offers a promising solution for precise diagnosis and management of retinal diseases.
• This work contributes to advancing AI applications in ophthalmology for better patient care.