Deep Learning with a Dataset Created Using Kanno Saitama Macro, a Self-Made Automatic Foveal Avascular Zone...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Foveal avascular zone (FAZ) extraction from optical coherence tomography angiography (OCTA) is crucial for diagnosing ophthalmic diseases.
• Manual FAZ extraction is time-consuming and subjective, posing challenges for large-scale deep learning applications.

Purpose of the Study:

• To evaluate the utility of a deep learning dataset for FAZ extraction generated by the Kanno Saitama Macro (KSM) program.
• To compare the performance of automated FAZ extraction using KSM and U-Net with manual extraction by expert examiners.

Main Methods:

• A dataset was created using KSM for automated FAZ extraction from swept-source OCTA images.
• A U-Net model was trained on this dataset.
• The Jaccard coefficient was used to compare the agreement between two manual examiners and between each examiner and the U-Net model.

Main Results:

• The Jaccard coefficients demonstrated high agreement between manual examiners (0.931) and between examiners and the U-Net model (0.951 and 0.933).
• The U-Net model achieved significantly higher agreement with examiners compared to inter-examiner agreement (p < 0.001).
• The KSM-generated dataset yielded automated FAZ extraction comparable or superior to manual methods.

Conclusions:

• The dataset generated by KSM is highly effective for training deep learning models for FAZ extraction.
• Automated FAZ extraction using KSM shows excellent performance, potentially reducing the annotation burden in deep learning for ophthalmology.