Microaneurysm detection using fully convolutional neural networks

Highlights

• An automatic method for detecting microaneurysms in fundus images is proposed.

• It uses a fully convolutional neural network with batch normalization and Dice loss.

• It requires only two processing stages.

• Shows how to transfer knowledge between datasets in microaneurysm domain.

• Produces better results than state-of-the-art methods.

Abstract

Backround and Objectives

Diabetic retinopathy is a microvascular complication of diabetes that can lead to sight loss if treated not early enough. Microaneurysms are the earliest clinical signs of diabetic retinopathy. This paper presents an automatic method for detecting microaneurysms in fundus photographies.

Methods

A novel patch-based fully convolutional neural network with batch normalization layers and Dice loss function is proposed. Compared to other methods that require up to five processing stages, it requires only three. Furthermore, to the best of the authors’ knowledge, this is the first paper that shows how to successfully transfer knowledge between datasets in the microaneurysm detection domain.

Results

The proposed method was evaluated using three publicly available and widely used datasets: E-Ophtha, DIARETDB1, and ROC. It achieved better results than state-of-the-art methods using the FROC metric. The proposed algorithm accomplished highest sensitivities for low false positive rates, which is particularly important for screening purposes.

Conclusions

Performance, simplicity, and robustness of the proposed method demonstrates its suitability for diabetic retinopathy screening applications.

Introduction

Diabetes affects one in eleven adults (over 400 million people worldwide) [1]. Diabetic retinopathy (DR) is a microvascular complication of diabetes which is the leading cause of vision loss in the working-age population [2]. One out of three diabetics has DR [3] and one in ten diabetic patients develops most vision-threatening form of DR [4]. Early detection of DR can prevent blindness in 90% of cases [5].

DR screening is manually performed by ophthalmologists and trained graders through a visual inspection of fundus photographs (FP). Unfortunately, the grading process is time-consuming, tedious, and error-prone with high inter-observer variability. Due to the rising number of DR patients worldwide (expected to exceed 640m by 2040 [1]) and their location (75% live in underdeveloped areas [6]) the development of computer-assisted diagnosis and automatic DR screening approaches are of the utmost importance.

Microaneurysms (MAs) are spherical swellings of the capillaries caused by weakening of the vascular walls; they appear as small round red dots. They are the earliest clinical sign of DR and continue to be present as the disease progresses. Consequently, automated detection of MAs can drastically reduce the screening workload. MA detection is a challenging task even for the human eye due to many factors including uneven image illumination, reflections, limited resolution and media opacity. The boundaries of MAs are not always well-defined and local contrast to the background is low, even in high-resolution images. Moreover, MAs may be confounded with visually similar anatomical structures such as haemorrhages, junctions in thin vessels, disconnected vessel segments, dark patches on vessels, background pigmentation patches and dust particles on the camera lense.

In general, the majority of MA detection methods consists of up to five stages: 1) Preprocessing, 2) MA candidate extraction, 3) Vessel removal, 4) Candidate feature extraction, and 5) Classification. The main goal of preprocessing is to remove noise, correct non-uniform illumination, and to improve contrast between the MAs and background. The MA candidate extraction stage uses a simple algorithm to identify a reasonably small set of locations with somewhat “lesion-like” appearance, attempting to identify all actual lesions together with many false positive regions. The vessel removal stage addresses the large number of false positives that may otherwise be produced by vessels. Next, hand-crafted features are extracted from candidate regions; this is the most labour-intensive and time-consuming part of the design stage. Finally, a classifier is trained to distinguish MAs from non-MAs based on the extracted features.

Baudoin et al. [7] introduced the first MA detection algorithm applied to fluorescein angiogram images. They employed a mathematical morphology based approach to remove vessels and applied a top-hat transformation with linear structuring elements to detect MAs. Several methods were built on this approach [8], however, since intravenous use of fluorescein can cause death in 1 in 222 000 cases [9], such methods are not suited for screening purposes. Walter et al. [10] also used a top-hat based method and automated thresholding to extract MA candidates. They extracted 15 features and applied kernel density estimation with variable bandwith for MA classification. In general, morphology-based approaches are sensitive to changes in size and shape of structuring elements which result in significant variations in MAs detection results. Zhang et al. [11] proposed a method based on dynamic thresholding and correlation coefficients of a multi-scale Gaussian template. They used 31 manually designed features based on intensity, shape and response of a Gaussian filter. Veiga et al. [12] presented an algorithm using Law texture features. Support Vector Machines (SVM) were used in a cascading manner: first SVM was used to extract MA candidates whereas the second SVM performed final MA classification. Haloi [13] used a vanilla convolutional network with 3 convolutional layers and 2 fully connected layers to detect MAs. Javidi et al. [8] proposed a technique which used 2D Morlet wavelet to find MA candidates. At the next stage, a discriminative dictionary learning approach was employed to distinguish MAs from other structures. Srivastava et al. [14] used Frangi-based filters that were manually designed to distinguish vessels from red lesions. Filters were applied to multiple sized image patches to extract features. Finally, these features were classified using a SVM.

Compared to the methods mentioned above, the proposed algorithm requires only three stages instead of five (preprocessing, patch extraction and classification). There is no need for MA candidate detection, vessel removal or feature extraction. Furthermore, the proposed method does not require manually hand-crafted features, it automatically learns the most discriminative features for MA detection. The vast majority of MA detection algorithms employ features based on MA shape, colour and texture. Unfortunately, many image modalities makes it virtually impossible to model them manually. To address this challenge, a Convolutional Neural Network (CNN) was used. CNNs have emerged as a powerful family of algorithms for solving computer vision tasks such as object detection [15], semantic segmentation [16] and image classification [17]. Compared with Haloi [13] method, the presented algorithm proposes a novel fully convolutional neural network (FCNN) architecture and transfers knowledge between MA datasets.

Training CNNs from scratch is not a trivial task, as they require large amounts of labelled data for training. In the MA detection domain, public datasets are small, scarce, and local lesion annotations on a per-pixel level are almost non-existent (to the best of authors knowledge, only one such dataset exists [18]). Moreover, the CNNs have vast capacity as learning models with millions of learnable parameters. As a result, they are very prone to overfitting and various convergence difficulties. Consequently, the initial values of a network’s weights have paramount importance in the learning process, especially for avoidance of local minima and saddle points.

To address these challenges, prior knowledge in the form of a network’s weights can be transferred between models that are later fine-tuned with new data. Azizpour et al. [19] showed that the success of knowledge transfer depends on the similarity between the training dataset of a CNN, and the dataset to which the knowledge is transferred. Given the limited availability of large medical datasets, research on transfer learning in medical imaging is largely focussed on transferring knowledge from general natural images datasets. However, these datasets have very different properties to medical datasets, including the fact that in medical datasets objects of interest may be very small and boundaries are of paramount importance. Consequently, knowledge transfer between these two domains is not optimal and produces various success rates [19]. In this paper we show that knowledge transfer even between small medical datasets can produce state-of-the-art results with an appropriate network architecture. To the best of our knowledge, this is the first time that deep transfer learning has been applied in the MA detection domain.

The main contributions of this paper are as follows. First, we propose a MA detection method that requires only three stages of analysis. Second, we present a novel CNN with a dedicated architecture for MA detection that does not require hand-crafted features. Third, we show how to successfully transfer knowledge between small datasets in MA domain - an important innovation in this domain as retinal image set characteristics vary between cameras, so that any practically useful method must be capable of simple and reliable retraining.

This paper is organized as follows. The proposed method is described in Section 2. Section 3 describes the datasets and performance metrics used for experiments. In Section 4 the evaluation results are presented and compared with existing approaches. Finally, in Section 5 discussion and conclusions are given.