Siamese and Recurrent neural networks for Medical Image Processing

In recent years computer vision applications have been pervaded by deep convolutional neural networks (CNNs). These networks allow practitioners to achieve the state of the art performance at least for the segmentation and classification of images and in object localization, but in each of these cases the obtained results are directly correlated with the size of the training set, the quality of the annotations, the network depth and the power of modern GPUs. The same rules apply to medical image analysis, although, in this case, collecting tagged images is more difficult than ever, due to the scarcity of data — because of privacy policies and acquisition difficulties — and to the need of experts in the field to make annotations. Very recently, scientific interest in the study and application of CNNs to medical imaging has grown significantly, opening up to challenging new tasks but also raising fundamental issues that are still open. Is there a way to use deep networks for image retrieval in a database to compare and analyze a new image? Are CNNs robust enough to be trusted by doctors? How can small institutions, with limited funds, manage expensive equipments, such as modern GPUs, needed to train very deep neural networks? This thesis investigates many of the issues described above, adopting two deep learning architectures, namely siamese networks and recurrent neural networks. We start with the use of siamese networks to build a Content–Based Image Retrieval system for prostate MRI, to provide radiologists with a tool for comparing multi–parametric MRI in order to facilitate a new diagnosis. Moreover, an investigation is proposed on the use of a composite loss classifier for prostate MRI, based on siamese networks, to increase robustness to random noise and adversarial attacks, yielding more reliable results. Finally, a new method for intra–procedural registration of prostatic MRIs based on siamese networks was developed. The use of recurrent neural networks is then explored for skin lesion classification and age estimation based on brain MRI. In particular, a new devised recurrent architecture, called C–FRPN, is employed for classifying natural images of nevis and melanomas allowing good performance with a reduced computational load. Similar conclusion can be drawn for the case brain MRI, where 3D images can be sliced and processed by recurrent architectures in an efficient though reliable way.