Attention-Guided 3D-CNN Framework for Glaucoma Detection and Structural-Functional Association Using Volumetric...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning (DL) models analyze 3D Optical Coherence Tomography (OCT) volumes for glaucoma biomarkers.
• Current state-of-the-art downsampling limits DL model learning from small retinal structures in OCT.
• Limited training data and computational resources necessitate efficient analysis of 3D OCT volumes.

Purpose of the Study:

• To improve DL model performance in learning finer ocular structures from 3D OCT volumes.
• To develop an end-to-end attention-guided 3D DL model for glaucoma detection and visual function estimation.

Main Methods:

• Proposed an attention-guided 3D DL model with three parallel pathways.
• Inputs included original 3D-OCT cubes and guided heatmaps generated during training.
• Model trained concurrently to minimize losses from all pathways, fusing predictions for the final output.

Main Results:

• Glaucoma detection model achieved 93.8% AUC, outperforming a baseline model (86.8%) and other ML approaches.
• Visual Field Index (VFI) estimation model achieved a Pearson correlation of 0.75 and median absolute error of 3.6%.
• The model demonstrated robustness and generalizability across classification and regression tasks.

Conclusions:

• The attention-guided 3D DL model significantly enhances glaucoma detection and visual function estimation from OCT scans.
• The approach effectively learns from fine retinal structures, overcoming limitations of traditional downsampling methods.
• This method offers a promising tool for objective glaucoma assessment and monitoring.