Performance of a deep learning-based convolutional neural network for quantification of outer retina atrophy compared...

The efficacy of candidate drugs for neuroprotection is tested in rodent models of retinal atrophy where light-induced outer retinal atrophy (ORA) is quantified by manual or semi-automated measurements, analyses that are time-consuming and error-prone. We developed a quantitative, automated image analysis-based method of ORA assessment in whole-slide images (WSIs). A commercial, cloud-based, artificial intelligence image analysis platform (Aiforia) was used to train convolutional neural network (CNN)-based deep learning models for semantic segmentation of retina and object counts of outer nuclear layer (ONL) nuclei. Model development was an iterative process of establishing and fine-tuning the ground truth (manual annotations), adding training annotations, and increasing the number of CNN training iterations. Inter-observer concordance yielded F1-scores exceeding 0.98 for retina area and 0.92 for ONL counts. Performance of the developed algorithm in comparison with manual annotations by individual validators yielded F1-scores between 0.89 and 0.95. Analysis of WSIs by the algorithm was completed 50× faster compared with manual analysis on hematoxylin and eosin-stained sections; however, it required time to digitize the slides. Performance of the ORA quantification deep learning model is noninferior to that of current manual ORA scoring approaches and provides improved reproducibility and reduced time and cost of analysis.