AutoMorph

AutoMorph automates extraction and quantification of retinal vascular morphology from fundus photographs using deep learning to support ophthalmic and systemic disease research.

Key Features:

• Functional Modules: Composed of four modules—Image Preprocessing, Image Quality Grading, Anatomical Segmentation, and Vascular Morphology Feature Measurement—implementing the full analysis workflow.
• Image Preprocessing: Prepares input fundus photographs for downstream analysis.
• Image Quality Grading: Uses the EfficientNet-b4 architecture for image quality assessment and achieves an F1-score of 0.86 on the EyePACS-Q dataset.
• Anatomical Segmentation: Performs binary vessel segmentation (F1 0.73 on AV-WIDE, 0.78 on DR HAGIS), artery/vein classification (F1 0.66 on IOSTAR-AV), and optic disc/cup segmentation (accuracy 0.94 on IDRID).
• Vascular Morphology Feature Measurement: Extracts detailed vascular features from segmented images with good to excellent agreement with expert annotations.
• Model Ensemble Strategy: Employs ensemble models across modules to improve robustness to domain differences and varying imaging devices.
• Confidence Analysis: Incorporates a confidence analysis mechanism to identify and rectify false gradable cases, reducing such errors by 76%.

Scientific Applications:

• Ophthalmology: Enables quantitative analysis of retinal vascular features for studies of ocular conditions.
• Oculomics: Supports oculomics research linking retinal vascular morphology to systemic and population-level health outcomes.
• Systemic disease research: Facilitates investigation of systemic diseases that manifest in the retina.

Methodology:

Uses deep learning including EfficientNet-b4, ensemble modeling, confidence analysis, anatomical segmentation (binary vessel, artery/vein, optic disc/cup), image quality grading, and image preprocessing, with validation on EyePACS-Q, AV-WIDE, DR HAGIS, IOSTAR-AV, and IDRID.