Classifying Retinal Degeneration in Histological Sections Using Deep Learning.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Histopathology

Background:

• Retinal degeneration classification is crucial for diagnosis and treatment.
• Artificial intelligence (AI) offers potential for automated image analysis in ophthalmology.
• Convolutional neural networks (CNNs) are powerful tools for image classification tasks.

Purpose of the Study:

• To evaluate the efficacy of a CNN in classifying histological images of retinal degeneration.
• To compare the performance of the CNN model against human observer classifications.
• To investigate the impact of reduced training data and image context on model performance.

Main Methods:

• A chemically induced feline model of monocular retinal dystrophy was utilized.
• Histological images were split into training and testing sets.
• Various CNN architectures were trained and the best performing model was evaluated against six human observers.

Main Results:

• The best CNN model achieved weighted-F1 scores between 85% and 90%.
• Cohen kappa scores reached 0.86, indicating high agreement between the model and observers.
• Reduced image context and training set size led to performance decreases of up to 6% and 10%, respectively.

Conclusions:

• Deep learning classifiers can reliably detect and grade retinal degeneration from histological data.
• AI models show promise in assisting the evaluation of complex retinal changes.
• This study provides a foundation for AI applications in analyzing clinical retinal image data.