Deep-Learning-Based Group Pointwise Spatial Mapping of Structure to Function in Glaucoma.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Understanding the spatial relationship between optic nerve head (ONH) structure and visual field (VF) function is crucial for diagnosing and managing glaucoma.
• Deep learning (DL) models show promise in predicting VF sensitivities from 3D Optical Coherence Tomography (OCT) scans.

Purpose of the Study:

• To establish generalizable pointwise spatial relationships between ONH structure and VF sensitivity using occlusion analysis of a DL model.
• To visualize and understand the contribution of specific ONH regions to VF predictions.

Main Methods:

• A DL model was trained on 12,915 3D OCT-VF pairs to predict 52 VF sensitivities.
• Occlusion analysis systematically evaluated the impact of individual ONH voxels on VF predictions in a test set of 996 pairs.
• Group t-statistic maps were generated to visualize statistically significant ONH regions corresponding to each VF test point.

Main Results:

• The study identified influential structural locations in the ONH for VF sensitivity prediction in both healthy-to-early-glaucoma and moderate-to-advanced-glaucoma groups.
• The revealed spatial correlations between OCT structure and VF function aligned with existing ophthalmic knowledge.
• Two-dimensional group t-statistic maps effectively assigned related ONH regions to specific VF test points.

Conclusions:

• This research successfully visualized point-by-point structure-function relationships in the ONH and VF without requiring prior segmentation or knowledge.
• The findings demonstrate the potential of machine learning models to provide robust, unbiased insights into glaucoma.
• The study opens possibilities for learning from trained ML models without pre-existing assumptions, enhancing diagnostic capabilities.