Automatic Segmentation of Retinoblastoma in Fundus Image Photography using Convolutional Neural Networks

Abstract

Purpose : Retinoblastoma is the most common form of ocular cancer in children. Locating newly-formed tumors in images of the retinal wall and estimating their growth pattern are significant challenges in the evaluation and tracking of the disease progression. In order to perform these tasks, ophthalmologists need to analyze images from different screening sessions and to annotate them, which is both time-consuming and tedious. Here, we present a method for the automatic segmentation of retinoblastoma in fundus, paving the way towards longitudinal tracking of ocular tumors.

Methods : Our data set is composed of 1186 fundus images with size 640 x 480 pixels from 28 pathological children eyes. Imaging was performed with a Retcam (Clarity Medical Systems, Pleasanton (CA), US) Fundus pediatric camera, with the patients under general anesthesia. The images were manually annotated by an expert ophthalmologist.
We achieve automatic segmentation of retinoblastoma by using a convolutional neural network (CNN) with an architecture known as U-Net (Ronneberger et al. in 2015). We trained the network with a reduced number of filters and use data augmentation techniques to enlarge the number of training samples (mirroring on the x and y axes). We furthermore modified the U-Net architecture with a seven-layer-deep structure with a geometrically increasing number of feature maps at every level.
We quantitatively evaluate our segmentation method with a 7-fold cross validation, testing on 4 of the 28 subjects every time and training on the rest (24). The resulting segmentation is assessed by the overlapping Dice Similarity Coefficient (DSC), with the manual segmentation used as the ground truth.

Results : Through the 7-fold cross-validation we obtain a DSC of 71.84±23.70% when using a smooth DSC as an objective function compared to a DSC of 62.45±19.63% when using binary cross-entropy. The testing time for the automatic segmentation is on average < 1 second, which, compared to manual delineation taking 1-2 minutes, results in a ~100x speed increase.

Conclusions : We have shown a fast segmentation tool that provides a significant benefit for the task of segmenting retinoblastoma in fundus images. In the future, this tool will be used to simplify longitudinal tracking of ocular tumors and the evaluation of competing treatment strategies.

This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.

 

Results of automatic segmentation of retinoblastoma in fundus.