advertisement

Editors Selection IGR 19-4

Clinical Examination Methods: Artificial Intelligence applications III

Comment by Andrew Tatham on:

82788 Deep Learning Approaches Predict Glaucomatous Visual Field Damage from OCT Optic Nerve Head En Face Images and Retinal Nerve Fiber Layer Thickness Maps, Christopher M; Bowd C; Belghith A et al., Ophthalmology, 2020; 127: 346-356

Find related abstracts

Deep learning (DL) is a machine learning technique that uses artificial neural networks inspired by the human brain, to learn from large amounts of data.¹ DL algorithms can automatically recognize intricate patterns in high-dimensional data and so are well suited for medical image analysis, including analysis of fundus photographs and optical coherence tomography (OCT).

In this innovative study, Christopher and colleagues developed DL models for detecting eyes with glaucomatous visual field (VF) loss and for predicting the severity of VF damage. Over 1,000 participants were included, with DL models trained to use RNFL thickness maps, RNFL en-face images or confocal scanning laser ophthalmoscopy (CSLO) images to predict the presence of VF loss and estimate VF indices including mean deviation (MD) and sectoral pattern deviation (PD).

The best performing DL model, which was based on the RNFL en-face image, achieved an area under the receiver operating characteristic curve (AUC) of 0.88 for detecting glaucomatous VF damage. This was significantly better than the performance of raw circumpapillary or global RNFL thickness values (AUC = 0.80 and 0.82 respectively). The RNFL en-face DL model also performed well in identifying eyes with early glaucomatous VF damage, defined as a MD > -6dB (AUC = 0.82 for the DL model versus 0.70 for circumpapillary RNFL thickness). In addition, the DL models were better at predicting MD, with the RNFL en-face model achieving a R² of 0.70 (mean absolute error of 2.5 dB) compared to a R² of only 0.45 for circumpapillary RNFL thickness (mean absolute error of 3.7 dB). The strongest sectoral associations for predicting VF PD from OCT were in the superior nasal and inferior nasal sectors of the VF.

This study shows that DL can also be used to directly model structure-function relationships and identify eyes with likely functional loss from OCT imaging

DL algorithms have recently been developed that have similar or better ability to detect glaucoma from color fundus images compared to ophthalmologists² and that can improve segmentation and identify neural and connective tissues of the optic nerve head.³ This study shows that DL can also be used to directly model structure-function relationships and identify eyes with likely functional loss from OCT imaging. The ability to better predict VF loss from OCT may help clinicians more effectively individualize the frequency of VF testing. For example, if the DL algorithm predicts VF loss similar to a patient's previous VF test, the clinician may decide to postpone repeat VF testing, whereas if the algorithm predicts worsening VF, a sooner repeat test may be indicated.

A further potential benefit is that when using OCT in clinical practice, clinicians often may place too much emphasis on the results of automated classifications or on global thickness measures. A major strength of the DL models is that they used information from the entire scan. In addition, DL models learnt about the structure-function relationship through training and did not make assumptions about linearity, a potential flaw of previous structure-function models.

Examining features of imaging, other than thickness, e.g., texture or voxel intensity, may provide additional value for glaucoma detection and monitoring.

The observation that DL models based on the RNFL en-face images outperformed models based on RNFL thickness maps and CSLO images is interesting. A possible explanation is that RNFL en-face images are computed from voxel intensity values within the RNFL, information not available from thickness values alone. This finding provides evidence that examining features of imaging, other than thickness, e.g., texture or voxel intensity, may provide additional value for glaucoma detection and monitoring.

Importantly, the authors also presented an intuitive method for clinicians to visualize the results of the DL model using heat maps to highlight regions of images of most importance in contributing to the classification process. Such a visualization technique could also prove useful in revealing fine-scale information about structure-function relationships.

• ← Previous
• Next →