Robust and Interpretable Convolutional Neural Networks to Detect Glaucoma in Optical Coherence Tomography Images

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning systems, particularly convolutional neural networks (CNNs), show promise in medical diagnosis, including ophthalmology for detecting retinal and ocular diseases.
• Clinical adoption of AI in medicine is hindered by concerns about algorithmic robustness across different datasets and a lack of explainability in AI predictions.
• Existing AI models often lack transparency, making it difficult to understand how they arrive at diagnostic conclusions compared to human experts.

Purpose of the Study:

• To develop robust CNN architectures for glaucoma detection using optical coherence tomography (OCT) images.
• To utilize concept activation vectors (TCAVs) to understand the image concepts CNNs use for predictions.
• To compare AI-derived concepts with human expert eye fixations to identify shared decision-making features.

Main Methods:

• Development of end-to-end deep learning models using fine-tuned transfer learning and CNN ensemble learning.
• Application of concept activation vectors (TCAVs) to interpret CNN predictions on OCT images.
• Comparison of TCAV results with eye-tracking data from clinicians examining OCT images.

Main Results:

• The developed deep learning models demonstrated superior robustness in glaucoma detection compared to previous hybrid models.
• TCAV analysis revealed that CNNs focus on specific OCT report sub-images for glaucoma detection.
• Comparison showed consistency between AI-identified concepts and areas fixated upon by expert clinicians.

Conclusions:

• Fine-tuned transfer learning and CNN ensemble learning enhance the robustness of AI models for glaucoma detection.
• TCAV analysis, validated by clinician eye-tracking, provides a method to interpret AI decisions and identify key diagnostic features.
• The proposed pipeline for evaluating AI robustness and explainability can facilitate the clinical acceptance of AI tools in ophthalmology.