Automatic segmentation of hyperreflective foci in OCT images

Highlights

• The leading cause of vision loss in the Western World is Age-related Macular Degeneration.

• The diagnosis of AMD is commonly done by the analyzing biomarkers on OCT images such as Hyperreflective Foci (HF).

• We proposed a method for training DNN-s for the automatic segmentation of HF.

• Automatic segmentation methods perform well on clinical data.

Abstract

Background and Objective

The leading cause of vision loss in the Western World is Age-related Macular Degeneration (AMD), but together with modern medicines, tracking the number of Hyperreflective Foci (HF) on Optical Coherence Tomography (OCT) images should assist the treatment of patients. Here, we developed a framework based on deep learning for the automatic segmentation of HF in OCT images.

Methods

We collected OCT images and annotated them, then these images underwent image preprocessing, and feature extraction steps. Using the prepared data we trained different types of Conventional-, Deep- and Convolutional Neural Networks to perform the task of the automatic segmentation of HF.

Results

We evaluated the various Neural Networks, by performing HF segmentation of clinical data belonging to patients, whose data were excluded from the training process. The results suggest that our systems can achieve reasonably high Dice Coefficient values, and they are comparable with (i.e., in most cases above 95%) the similarity between manual annotations performed by different physicians.

Conclusion

From the results, it can be concluded that neural networks can be used to accurately segment HF in OCT images. The results are sufficiently accurate for us to incorporate them into the next phase of the research, building a decision support system for everyday clinical practice.

Introduction

The importance of artificial intelligence (AI) and Deep learning have already been proven in many medical fields such as radiology, pathology, and dermatology. With the development of Machine Learning (ML), a branch of AI, it is now possible to automatically recognize anatomical structures or lesions in medical images (segmentation); to place an image into different categories (classification); and also to predict the outcome of a process (prediction). In ophthalmology, thanks to 21st century diagnostic innovations, the Optical Coherence Tomography (OCT), retinal structures can be visualized with an unprecedented resolution. With these high resolution OCT scans, we can now develop deep learning methods and utilize machine learning algorithms in order to reduce diagnostic and therapeutic errors, and promote personalized medicine [1]. ML has already provided clinically acceptable diagnostic performance in detecting many retinal diseases, such as diabetic retinopathy (DR), glaucoma, retinopathy of prematurity (ROP) and Age-related Macular Degeneration (AMD) [2].

The global increase in life expectancy has led to an increase in the number of age-related diseases, and AMD has become the leading cause of vision loss in the Western World [3], and a health problem worldwide [4]. Because of the large prevalence of AMD, the proper management of the disease is crucial.

With the help of OCT, several disease relevant biomarkers have been recently identified in AMD, such as Hyperreflective Foci (HF), which could provide a sensitive marker for the treatment decision process. However their manual quantification is quite challenging [5], [6]. Therefore, we addressed the task in an interdisciplinary fashion, combining the expertise of clinicians with modern computational tools taken from the fields of image processing and artificial intelligence.

In our study, a framework was developed where the ophthalmologists could mark HF on OCT scans. Then, the marked images were transformed using various image processing techniques in order to extract features which best characterise HF. Quite recently, Deep Neural Networks have achieved excellent results in medical data analysis [7], [8], [9], and in the next step, image data sets were used for training Artificial Neural Networks (ANN), Deep Neural Network (DNN), and Convolutional Neural Networks (CNN). These networks were later used in the automatic detection of HF. Our methods were also validated on clinical data, gathered from patients of the Department of Ophthalmology at the Clinical Center at the University of Szeged.

A full summary of related scientific work can be found in Section 2. Previous research on the topic mainly focused on the automated evaluation of HF in manually annotated datasets [10], [11], [12], [13], [14], [15] without automated evaluation, while our study focuses on the automatic segmentation of HF. One parallel study on this topic is available [16] on arxiv, where the authors applied deep learning-based methods using variants of the U-Net [9] for the automatic segmentation of HF. The approach described by our group also differs from the previous study involving training nets from two parallel sets of uncertain data. Moreover, we performed two independent evaluations of our models by comparing the network outputs with the annotations of two clinical experts.

In another study, the authors of [17] applied similar methods to ours for the quantization of HF in OCT images, but they did not perform a pixel-wise evaluation of the results. In their study, they used a smaller test dataset, and now we propose new methods that perform better.

The novelty of our study lies in testing various neural networks for the automatic segmentation of HF. Unlike previous studies, neural networks were trained not by only relying on raw pixel information, but also on features extracted by image preprocessing steps. We compared the annotations of two different clinical doctors with each other to determine a baseline accuracy for HF segmentation as a reference for our methods. We evaluated our methods on two test datasets, and we found that our methods led to Dice Coefficients values, which are comparable with the similarity score between manual annotations performed by physicians. Lastly, we demonstrated that in contrast to current trends, small networks can also provide reasonably good results, and they do not need enormous amounts of training. Our given methods are robust in the sense that they were trained using a small amount of uncertain data, and this allows one to adapt them to new environments with a minimal burden on the valuable time of medical staff.