Multiple model visual feature embedding and selection method for an efficient oncular disease classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Early detection of ocular diseases is crucial for preventing vision loss but faces challenges in specialist availability and resource limitations.
• Automated diagnostic tools are needed to improve precision and streamline clinical workflows for ocular disease identification.

Purpose of the Study:

• To develop and evaluate a deep learning-based system for automated classification of ocular diseases from fundus images.
• To optimize computational efficiency and diagnostic performance using a novel feature selection framework.

Main Methods:

• Utilized the Ocular Disease Intelligent Recognition (ODIR) dataset comprising 5,000 labeled fundus images.
• Employed transfer learning models (DenseNet201, EfficientNetB3, InceptionResNetV2) and a two-level feature selection framework combining Linear Discriminant Analysis (LDA) with advanced neural networks (DNN, LSTM, BiLSTM).

Main Results:

• The combined feature strategy with Bidirectional LSTM (BiLSTM) achieved 100% accuracy, precision, and recall on the training set.
• Over 98% performance was recorded on the validation set, demonstrating high efficacy.
• The LDA-based framework significantly reduced computational complexity while improving classification accuracy.

Conclusions:

• The proposed deep learning system offers a scalable and efficient solution for automated ocular disease detection.
• This technology provides robust support for clinical decision-making, potentially reducing ophthalmologist workload, especially in resource-limited areas.
• The system has the potential to improve global patient outcomes through enhanced early detection of eye conditions.