Deep Learning-Based Cataract Detection and Grading from Slit-Lamp and Retro-Illumination Photographs: Model...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Cataract diagnosis and grading are crucial for effective treatment.
• Current methods may face limitations in accuracy and consistency.
• Automated systems offer potential for improved efficiency and standardization.

Purpose of the Study:

• To develop and validate a deep learning (DL)-based artificial intelligence (AI) platform for cataract diagnosis and grading.
• Utilize slit-lamp and retroillumination lens photographs based on the Lens Opacities Classification System (LOCS) III.
• Overcome limitations of medical databases, such as small dataset size and biased label distribution.

Main Methods:

• A convolutional neural network was trained and tested on 1335 slit-lamp and 637 retroillumination lens images from 596 patients.
• Employed strategies including region detection, data augmentation, transfer learning, generalized cross-entropy loss, and class-balanced loss.
• Ensemble of ResNet18, WideResNet50-2, and ResNext50 algorithms reinforced AI performance.

Main Results:

• The AI platform achieved robust diagnostic performance for nuclear opalescence (NO) and nuclear color (NC) with high AUC and accuracy.
• Demonstrated high diagnostic performance for cortical opacity (CO) and posterior subcapsular opacity (PSC).
• Achieved good LOCS III-based grading prediction performance for all cataract types.

Conclusions:

• The developed DL-based AI platform accurately detects and grades various cataract types (NO, NC, CO, PSC).
• The system effectively addresses challenges posed by limited training data and biased label distribution in medical datasets.
• This AI platform shows significant promise for automated cataract assessment in clinical practice.