Automatic Cataract Classification Using Deep Neural Network With Discrete State Transition.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Cataract is the primary cause of global blindness, necessitating improved detection and grading methods.
• Accurate and convenient automated systems are crucial for managing vision impairment.
• Current methods require enhancement for precision and efficiency.

Purpose of the Study:

• To develop and evaluate automatic methods for cataract detection and severity grading.
• To compare the performance of different feature extraction and classification strategies.
• To introduce novel deep learning approaches for medical image analysis.

Main Methods:

• Proposed methods include Haar and visible structure features with discrete state transition multilayer perceptron (DST-MLP) or exponential DST (EDST-MLP) classifiers, utilizing prior knowledge.
• Developed residual neural networks with DST (DST-ResNet) or EDST (EDST-ResNet) for classification without prior knowledge.
• Employed DST and EDST strategies to mitigate overfitting and reduce memory usage.

Main Results:

• Combined features outperformed single features in cataract detection and grading accuracy.
• Classification methods incorporating prior knowledge-based feature extraction demonstrated superior performance.
• The proposed DST and EDST strategies achieved state-of-the-art accuracy while optimizing network training and implementation.

Conclusions:

• Automated cataract detection and grading using AI, particularly with combined features and prior knowledge, significantly enhances diagnostic capabilities.
• The DST and EDST strategies offer effective solutions for preventing overfitting and reducing memory requirements in neural network applications.
• These findings provide valuable insights for advancing medical image processing and AI in healthcare.