Deep learning-based image quality assessment for optical coherence tomography macular scans: a multicentre study.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate assessment of optical coherence tomography (OCT) image quality is crucial for reliable diagnosis.
• Deep learning (DL) offers potential for automating image quality assessment in retinal imaging.

Purpose of the Study:

• To develop and externally validate deep learning models for assessing the image quality of 3D macular scans from Cirrus and Spectralis OCT devices.
• To evaluate the performance of these DL models on unseen datasets from diverse international sources.

Main Methods:

• Retrospective collection of 2277 Cirrus and 1557 Spectralis 3D OCT scans for training, fine-tuning, and internal validation.
• Utilized a 3D Residual Network (ResNet)-18 for Cirrus and a multiple-instance learning pipeline with ResNet-18 for Spectralis.
• External testing involved three unseen Cirrus datasets and five unseen Spectralis datasets from multiple countries.

Main Results:

• Internal validation demonstrated high performance with Area Under the Curve (AUC) of 0.930 for Cirrus and 0.906 for Spectralis scans.
• External testing confirmed robust performance, with AUCs for Cirrus ranging from 0.832 to 0.930 and for Spectralis from 0.891 to 0.962.

Conclusions:

• The developed DL models can effectively identify and filter out ungradable 3D OCT macular scans.
• Integration of these models into a disease-detection workflow can facilitate fully automated eye disease detection.