Clinically applicable deep learning for diagnosis and referral in retinal disease

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Increasing volume and complexity of diagnostic imaging outpace human expertise.
• Current AI models for 2D images require millions of annotated images.
• AI performance in 3D diagnostic scans within real-world clinical pathways remains a challenge.

Purpose of the Study:

• To develop and validate a novel deep learning architecture for interpreting 3D optical coherence tomography (OCT) scans.
• To achieve expert-level performance in making referral recommendations for sight-threatening retinal diseases.
• To demonstrate the device-independent nature of the AI's tissue segmentation for broader clinical applicability.

Main Methods:

• Application of a novel deep learning architecture to a heterogeneous set of 3D OCT scans.
• Training the model on 14,884 annotated OCT scans.
• Evaluating the model's referral recommendation accuracy and tissue segmentation capabilities.

Main Results:

• The deep learning model achieved referral recommendation performance matching or exceeding expert clinicians.
• The model was trained on a relatively small dataset (14,884 scans) compared to traditional AI methods.
• Tissue segmentations generated by the AI were device-independent, maintaining accuracy across different OCT devices.

Conclusions:

• The novel deep learning architecture effectively interprets 3D OCT scans for diagnosing retinal diseases.
• This approach overcomes the need for massive annotated datasets, facilitating wider clinical adoption.
• The AI's performance and device independence offer a promising solution for managing the growing demand for diagnostic imaging expertise.