Purchase this article with an account.
Alessia Colonna, Fabio Scarpa, Riccardo Zorzan, Cecilia Chao, Rabia Mobeen, Blanka Golebiowski, Fiona Stapleton, Alfredo Ruggeri; Automatic identification of corneal epithelial dendritic cells via deep learning. Invest. Ophthalmol. Vis. Sci. 2020;61(7):4309.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
In-vivo confocal microscopy (IVCM) has been extensively used to image all layers and structures of the cornea. Recent studies investigated the role of corneal epithelial dendritic cells (CEDCs) in some ocular pathologies, but identifying and tracing of CEDCs using manual methods is time consuming and subjective. The aim of this study is to develop an automated reliable algorithm, based on the Deep Learning technique Convolutional Neural Network, which is able to detect and measure CEDCs.
100 confocal images, a field of 400x400 μm (384x384 pixels), of the sub-basal corneal nerve plexus were randomly selected from images captured at the central and mid-peripheral (temporal and superior) cornea of 20 healthy and 20 post-LASIK participants. Images were acquired with the Heidelberg Retina Tomograph (HRT-II) with the Rostock Cornea Module (Heidelberg Engineering GmbH, Heidelberg, Germany) at UNSW, Sydney.In each image, the contour of each CEDC was traced manually. The tracing of nerve fibers was also included in the ground truth during the training process. Nerves were automatically traced using an algorithm previously developed by our group (Guimarāes et al. TVST, 2016). A data augmentation technique (flipping each image horizontally and vertically) was used to obtain a larger dataset (300 images in total) for training process. An U-shaped Convolutional Neural Network (U-Net) was then implemented for both a contracting-encoder path that provides useful features for CEDC identification and an expanding-decoder path that locates CEDC in the images (FIG.1.).The U-Net was trained on 297 images (99 original+198 flipped) and evaluated on the remaining original image, and this was repeated in turn for the set of 100 original images (leave one out cross-validation).
Manual and automated segmentation were compared at a single-pixel level and at a cellular level (region of connected pixels). For pixels, the True Positive Rate (percentage of CEDC correctly traced) was 92.0%, and the False Discovery Rate (percentage of CEDC wrongly traced) was 37.6%. For regions, the True Positive Rate was 96.3%, and the False Discovery Rate was 20.0% (FIG.2).
The proposed method appears capable of extracting features that correctly describe dendritic cells, allowing their accurate identification. These results encourage the further development of the proposed method and its application in clinical investigations.
This is a 2020 ARVO Annual Meeting abstract.
This PDF is available to Subscribers Only