Classifying Retinal Degeneration in Histological Sections Using Deep Learning.

Purpose: Artificial intelligence (AI) techniques are increasingly being used to classify retinal diseases. In this study we investigated the ability of a convolutional neural network (CNN) in categorizing histological images into different classes of retinal degeneration. Methods: Images were obtained from a chemically induced feline model of monocular retinal dystrophy and split into training and testing sets. The training set was graded for the level of retinal degeneration and used to train various CNN architectures. The testing set was evaluated through the best architecture and graded by six observers. Comparisons between model and observer classifications, and interobserver variability were measured. Finally, the effects of using less training images or images containing half the presentable context were investigated. Results: The best model gave weighted-F1 scores in the range 85% to 90%. Cohen kappa scores reached up to 0.86, indicating high agreement between the model and observers. Interobserver variability was consistent with the model-observer variability in the model's ability to match predictions with the observers. Image context restriction resulted in model performance reduction by up to 6% and at least one training set size resulted in a model performance reduction of 10% compared to the original size. Conclusions: Detecting the presence and severity of up to three classes of retinal degeneration in histological data can be reliably achieved with a deep learning classifier. Translational Relevance: This work lays the foundations for future AI models which could aid in the evaluation of more intricate changes occurring in retinal degeneration, particularly in other types of clinically derived image data.

Electrical Field Shaping Techniques in a Feline Model of Retinal Degeneration.

The majority of preclinical studies investigating multi-electrode field shaping stimulation strategies for retinal prostheses, have been conducted in normally-sighted animals. This study aimed to reassess the effectiveness of two electrical field shaping techniques that have been shown to work in healthy retinae, in a more clinically relevant animal model of photoreceptor degeneration. Four cats were unilaterally blinded via intravitreal injections of adenosine triphosphate. Cortical responses to traditional monopolar (MP) stimulation, focused multipolar (FMP) stimulation and two-dimensional current steering were recorded. Contrary to our previous work, we found no significant difference between the spread of cortical activation elicited by FMP and MP stimulation, and we were not able to reproduce cortical responses to singleelectrode retinal stimulation using two-dimensional current steering. These findings suggest that while shown to be effective in normally-sighted animals, these techniques may not be readily translatable to patients with retinal degeneration and require further optimization.

Creating virtual electrodes with 2D current steering.

OBJECTIVE: Current steering techniques have shown promise in retinal prostheses as a way to increase the number of distinct percepts elicitable without increasing the number of implanted electrodes. Previously, it has been shown that 'virtual' electrodes can be created between simultaneously stimulated electrode pairs, producing unique cortical response patterns. This study investigated whether virtual electrodes could be created using 2D current steering, and whether these virtual electrodes can produce cortical responses with predictable spatial characteristics. APPROACH: Normally-sighted eyes of seven adult anaesthetised cats were implanted with a 42-channel electrode array in the suprachoroidal space and multi-unit neural activity was recorded from the visual cortex. Stimuli were delivered to individual physical electrodes, or electrodes grouped into triangular, rectangular, and hexagonal arrangements. Varying proportions of charge were applied to each electrode in a group to 'steer' current and create virtual electrodes. The centroids of cortical responses to stimulation of virtual electrodes were compared to those evoked by stimulation of single physical electrodes. MAIN RESULTS: Responses to stimulation of groups of up to six electrodes with equal ratios of charge on each electrode resulted in cortical activation patterns that were similar to those elicited by the central physical electrode (centroids: RM ANOVA on ranks, p > 0.05; neural spread: one-way ANOVA on Ranks, p > 0.05). We were also able to steer the centroid of activation towards the direction of any of the electrodes of the group by applying a greater charge to that electrode, but the movement in the centroid was not found to be significant. SIGNIFICANCE: The results suggest that current steering is possible in two dimensions between up to at least six electrodes, indicating it may be possible to increase the number of percepts in patients without increasing the number of physical electrodes. Being able to reproduce spatial characteristics of responses to individual physical electrodes suggests that this technique could also be used to compensate for faulty electrodes.

Spatial Restriction of Neural Activation Using Focused Multipolar Stimulation With a Retinal Prosthesis.

PURPOSE: The resolution provided by present state-of-the-art retinal prostheses is severely limiting for recipients, partly due to the broad spread of activation in the retina in response to monopolar (MP) electrical stimulation. Focused multipolar (FMP) stimulation has been shown to restrict neural activation in the cochlea compared to MP stimulation. We extended the FMP stimulation technique to a two-dimensional electrode array and compared its efficacy to MP and hexapolar (HP) stimulation in the retina. METHODS: Normally-sighted cats (n = 6) were implanted with a suprachoroidal electrode array containing 42 electrodes. Multichannel multiunit spiking activity was recorded from the visual cortex in response to MP, HP, and FMP retinal stimulation. RESULTS: When inferring retinal spread using voltage recordings off the stimulating array, FMP stimulation showed significantly reduced voltages in regions surrounding the primary stimulating electrode. When measuring the retinal and cortical selectivity of neural responses, FMP and HP stimulation showed significantly higher selectivity compared to MP stimulation (separate 2-way ANOVAs, P < 0.05). However, the lowest cortical thresholds for each stimulating electrode were higher for FMP and HP compared to MP stimulation (1-way ANOVA, P < 0.001). No significant differences were observed between FMP and HP stimulation in any measures. CONCLUSIONS: Focused multipolar and HP stimulation using a two-dimensional array are promising techniques to reduce the spread of activation for a retinal prosthesis. Clinical application would be expected to result in smaller phosphenes; thus, reducing phosphene overlap between electrodes and increasing the resolution at the expense of higher thresholds for activation.