Abstract
Purpose :
Under hyperglycemic conditions in diabetes, degradation of the microvasculature occurs in the retina. Thus, leading to shift in the ratio of pericytes to endothelial cells, the two cell types that make up capillaries, which is 1:1 in the normal human retina. However, due to pericyte loss, the ratio becomes 1:4 in the diabetic human retina. In the field of diabetic retinopathy, researchers observe this ratio of pericyte and endothelial cells by manually counting each cell type. Therefore, we developed a machine learning model that will automatically classify these cell types and thus greatly reduce analysis time.
Methods :
Elastase digests were performed on mouse retinas to isolate the vasculature, which was stained with hematoxylin and Periodic Acid Schiff and imaged using light microscopy. Using CellProfiler we developed a pipeline to identify microvascular cells present in the elastase digest images. Cells at the image perimeter or overlapping were removed by our pipeline from analysis. Four different machine learning models were developed to classify cells as pericytes or endothelial cells, and the GridSearchCV technique was used to narrow down the hyperparameters by model. The metrics used to compare the performance of the models were the training set (n=2324 instances) and testing set (n=451 instances) accuracy. Manual (n=3) and automated image analysis were compared for time and cell identification (n=10).
Results :
After completing the development and testing for each machine learning model, the random forest classifier was chosen as the final model for analysis. The final model achieved an accuracy of approximately 93.80% on the training set and 91.57% on the testing set. Our automated program significantly reduced analysis time (2.57 minutes, n=10) to classify retinal vascular cells compared to manual analysis (54.28 minutes, n=10).
Conclusions :
These results give hope towards eliminating the need for manual classification and introducing a new way to classify these cell types more efficiently. Additionally, these machine learning models can be leveraged in other areas of diabetic retinopathy and help build towards the goal of better understanding this condition and reducing the risks it poses.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.