Identifying Age-Related Macular Degeneration with AI Deep Learning Model

Arnt-Ole Tvenning

An investigation aimed to examine which areas could be associated with age-related macular degeneration (AMD) with visualization of a deep learning model's point of interest (POI). The sudy used OptiNet, a deep learning model that was trained to identify the presence of AMD in spectral-domain optical coherence tomography (SD-OCT) macular scans.

SD-OCT has become a cornerstone of evaluation in ophthalmology, particularly when it comes to diagnosing AMD. As a result, the amount of images being collected is continuously increasing. Arnt-Ole Tvenning, Department of Ophthalmology, St. Olav Hospital, Trondheim University Hospital, and investigators sought to tap into the wealth of information these images have the potential to provide.

The OptiNet deep learning artificial intelligence (AI) tool was utilized because its artificial neural networks were structured with multiple layers of neurons to understand complex tasks in a similar way to the brain.

"Deep learning models learn without a priori knowledge, their own features of a given disease, and these features may then be used to classify medical images such as SD-OCT as either pathologicalor normal," investigators wrote.

Points of Interest of Deep Learning AI

OptiNet was trained and validated on 2 datasets for this research. The first set included a single scan of 269 cases of AMD and 115 controls. The second set consisted of 337 scans taken from 40 AMD cases (62 eyes) and 46 scans of both eyes of the 23 control cases.

Visualization of points of interests was achieved by calculating feature dependencies across the layer hierarchy in the deep learning architecture.

Patients were 50 years or older in the first dataset, and had intermediate AMD. Controls were age-matched and showed no signs of AMD in either eye. The second dataset involved patients of the same age group who had scans performed every 6 months.

The deep learning model was trained to extract features according to specific information from the scan in order to create a classification of data from the scans. Each convolution layer learned to identify relevant features, like lines and edges as well as the probability of AMD presence.

Classifying Data From SD-OCT Images

Investigators identified retinal nerve fiber (82%) and choroid layers (70%) as points of interest in cases that were classified as AMD. The areas applied most frequently were retinal pigment epithelium (98%) and drusen (97%).

OptiNet obtained area under the receiver operator curves of≥99.7%.

The results from the deep learning model indicated alterations in the SD-OCT imaging regions attributed to the retinal nerve fiber and choroid layers. More investigation of the neuroretina and choroid role in AMD development is needed to futher understand if the findings represent a change in macular tissue with AMD, according to investigators.

"For the first time, the anatomical areas utilized by a deep learning model to identify AMD are shown with a novel visualization method. This method revealed that regions such as the RNFL and choroid might be altered in AMD and demonstrate the potential of deep learning as a method not only in identification but also in the exploration of retinal disease," they wrote. "As a result, the methods applied in this study may be expanded to include multiple retinal diseases."

The study "Deep learning identify retinal nerve fibre and choroidlayers as markers of age-related maculardegeneration in the classification of macular spectral-domain optical coherence tomography volumes" was published in Acta Ophthalmologica.