Article Preview
Top1. Introduction
Recently, there has been an increasing interest in deep learning (DL) techniques within various fields: automatic machine translation (Klei et al., 2017), automatic text generation (He, & Deng, 2017), self-driving cars (Hadsell et al., 2009), image recognition (LeCun, 2015) and healthcare (Litjens et al., 2017).
In healthcare, computer-aided diagnostic systems (CAD) are employed to reduce the inter variability between experts. A good example of this was presented by (Elmore et al., 2015), where they illustrated that the concordance rate between pathologists' interpretations was 75.3%. Whereas, the exploitation of CAD systems could improve the detection of breast cancer by 10% (Gromet, 2008).
Previously, CAD systems were based on expert systems (Shortliffe, 1986) and machine learning (ML) methods. ML algorithms have attracted considerable attention for medical image analysis (Litjens et al., 2017). Nevertheless, these strategies rely on data representation, where the unsupervised extraction process of handcrafted features can reduce the classification precision when including indiscriminate features. The important obtained error rate in the ImageNet large scale visual recognition competition (ILSVRC) (Krizhevsky et al., 2012), and the availability of powerful graphical process units (GPU’s) (Chetlur et al., 2014) encouraged the image analysis community to take advantage of deep learning techniques in CAD systems. In computer vision, convolutional neural networks (CNNs) have attracted considerable attention compared to other DL algorithms. The particularity of these methods compared to traditional ML algorithms is their ability to perform supervised feature extraction by convolutional layers. Hence, this automation reduces the expert's workload and enables the direct application of these algorithms in various domains. Several CNN architectures have been proposed in the literature to solve different issues: AlexNet (Krizhevsky et al., 2012), VGGNET (Simonyan, & Zisserman, 2014), Inception (Szegedy et al., 2015), Xception (Chollet, 2017), Inception-Resnet (Szegedy et al., 2017), ResNet (He et al., 2016), ShuffleNet (Zhang et al., 2017), DenseNet (Huang et al., 2017) and MobileNet (Howard et al., 2017; Sandler et al., 2018). These architectures have received increased attention in imaging competitions (2D segmentation of brain electron microscopic (EM) images (Arganda-Carreras et al., 2015), multi-column max-pooling convolutional neural networks (MCMPCNN) for mitosis detection (Veta et al., 2015)) and several medical applications (mammography (Dubrovina et al., 2018), cardiovascular (Arganda-Carreras et al., 2015) and histopathology (Janowczyk, & Madabhushi, 2016).