Understanding Deep Convolutional Networks for Biomedical Imaging: A Practical Tutorial

Medical imaging seeks to unveil the internal structures hidden by the skin and bones to assist disease diagnosis and also treatment optimisation. In the past, processing medical images used to be a laborious task. However, the development of artificial intelligence has allowed the machine to gain a high level of understanding to perceive and extract information from biomedical images. Deep learning models, in particular, the convolutional neural networks (CNNs), were developed and implemented successfully for various biomedical applications. Therefore, it is of paramount importance for healthcare practitioners to understand the mechanisms behind the implemented CNNs to accurately interpret their outcomes. This tutorial summarises the key steps to train a functional CNNs. CNNs are usually constructed in the order of a convolution operation, ReLU, spatial pooling and followed by the fully connected layers. In addition, we have also introduced a number of preprocessing methods that target the image augmentation to combat the sparse data problem. We further explored a generative model as an augmentation method known as the generative adversarial networks (GANs), where GANs may yield new useful information to the dataset as compared to the classical augmentation.