Abstract
Purpose :
To develop a deep learning-based method with limited training resources, that can automatically identify and count the number of Retinal Pigment Epithelium (RPE) cells in confocal microscopy images obtained from cell culture or mice RPE/Choroid flat-mounts.
Methods :
The training and testing dataset contains two different image types: wild-type mice RPE/Choroid flat-mounts and ARPE 19 cells in culture, stained for Rhodamine-phalloidin, and imaged with confocal microscopy. Our approach is shown in Fig.1. Afterimage pre-processing for de-noising and contrast adjustment, Scale-Invariant Feature Transform descriptors are used for feature extraction. The set of training labels is derived from cells in original training images, annotated and converted to Gaussian density maps as a sum of fixed-variance Gaussians centered at each annotation. Finally, we train a Deep Neural Network (DNN) using the set of training input features and labels such that the obtained DNN model accurately predicts the corresponding Gaussian density maps, and thus identify/count the cells for any image.
Results :
The training and testing dataset contains 229 images from ARPE19 and 85 images from RPE/Choroid flat-mounts. Within two data sets, 30% and 10% of the images were selected for validation. We achieved 96.48 ± 6.56% and 96.88 ± 3.68% accuracy, with 95% confidence intervals, on ARPE19 and RPE/Choroid flat-mounts samples. Our approach outperforms existing relevant methods by significantly decreasing prediction error and variance, as shown in Fig. 2.
Conclusions :
We developed a deep learning-based approach that can accurately estimate the number of RPE cells contained in images obtained from cell culture and mice flat-mounts. We devised image pre-processing and data augmentation methods to form sufficient training images and improve the accuracy of the learning algorithm. Unlike existing methods that either requires a substantial number of training images or convex loss functions, our method achieves high accuracy with limited training datasets without compromising the expressiveness of the learning model. Furthermore, our approach is not limited by the cell shape of the input microscopy images and can be effectively used on images with unclear and curvy boundaries.
This is a 2020 ARVO Annual Meeting abstract.