advertisement

Editors Selection IGR 23-4

Artificial Intelligence: Structure-Function Correlation in the AI Era

Comment by Michael Girard & Thanadet Chuangsuwanich on:

113396 Prediction of Central Visual Field Measures From Macular OCT Volume Scans With Deep Learning, Mohammadzadeh V; Vepa A; Vepa A; Li C et al., Translational vision science & technology, 2023; 12: 5

Find related abstracts

Understanding the relationship between functional and structural measures in primary open-angle glaucoma is essential for grading the severity of the disease and assessing its natural progression. Numerous studies have reported strong correlations between optic nerve head structural changes (e.g., RNFL and neuroretinal rim thinning) and macula thinning with visual field impairment. However, significant challenges remain, as we often observe glaucomatous structural changes without detectable visual field abnormalities, and vice versa, in clinical settings. Additionally, current methods for determining and tracking visual field loss are prone to random errors and rely heavily on the individual's ability to concentrate, affecting the reliability of the results. Therefore, it is important to enhance our ability to infer glaucoma status from structural information, as this holds promise for improving early detection, monitoring, and treatment of the disease.

With advancements in artificial intelligence (AI), numerous studies have employed deep learning networks to predict visual field status from structural imaging. In this study, the authors utilized three trained convolutional neural networks (CNNs) on volumetric scans of the macula to predict mean deviation (MD), threshold sensitivity (TS), and total deviations (TD), performing a regression task for each parameter from the structural images. They compared the performance of the 3D CNNs with (1) 2D CNNs that use 2D tissue thickness maps as input; and (2) a standard regression model (baseline model) that uses ganglion cell/ inner plexiform layer thickness as input. The results showed that the 3D CNNs significantly outperformed the other two models in all tasks.

Overall, this study is particularly interesting due to its use of the full 3D volume of the macula and the large sample size of 1,121 subjects. The finding that using 3D volume as input to CNNs yields better prediction performance than using 2D tissue thickness maps alone underscores the importance of leveraging the 3D macular scan data. Additionally, the authors create an 'occlusion map', a technique used to understand how CNNs make decisions by highlighting which parts of an image are important for their predictions. This map is crucial as it reveals the structure-function relationship as interpreted by the CNNs within a specific dataset.

The finding that using 3D volume as input to CNNs yields better prediction performance than using 2D tissue thickness maps alone underscores the importance of leveraging the 3D macular scan data

Overall, the manuscript holds value in investigating the structure-function relationship in glaucoma. Its key messages are: (1) recent techniques like AI can be effectively used to study this relationship; and (2) comprehensive 3D information of the macula is strongly associated with functional loss. However, there are several areas where the manuscript could be improved. The discussion on how the findings, such as those from the occlusion map, could reveal new insights about the structure-function relationship is somewhat lacking. Additionally, the authors used a generic CNN, which may not be entirely suitable for the task, given the unique characteristics of macular volumes, such as the significant amount of empty area (without tissues). This current model processes both the empty area and the tissue, which may not be optimal. An alternative network or a more customized CNN that can better represent the macular structure and handle its low signal-to-noise ratio would be more appropriate.

• ← Previous
• Next →