Integration of Deep Convolutional Networks and Active Shape Models for automatic prostate segmentation

Prostate cancer is the most common cancer in the world for what concerns the male population. There are several types of tests to diagnose it, including screening tests based on the recognition of the PSA antigen, urological approaches with ultrasounds or other techniques that make use of different acquisition modalities. Among these, MRI proves to have the greatest accuracy in recognizing the prostate and tissue irregularities within it. The diagnosis of a pathology affecting this organ is carried out using CAD systems that use automatic segmentation algorithms as a support to recognize at first the organ in its entirety or a region in which it is most likely to be found within the image. The use of automatic algorithms for prostate segmentation allows to bypass the huge workload on technicians as the data available for processing is enormous and it needs rapid and less operator-dependent techniques. The latest approaches proposed use deep neural networks that rely on the large acquired capacity of computers in terms of GPU and CPU to automatically extract deep features from large datasets and to easily classify the input images on the basis of the repetition of recurring patterns. The state of the art regarding automatic prostate segmentation will be exposed and a series of methods proposed for this scope will be presented in detail. The anatomy of this organ will be outlined focalizing on its characteristics in MRI images to better understand the major difficulties in its segmentation, such as the less defined edges at the base and the apex, the small effective area in the whole MRI volume, the appearance of the organ similar to the one of all the tissues surrounding it, the large variability in shape among different patients and other variable characteristics that depend on the scanning protocols and noise corrupting the acquired images. The method we propose is fully automatic and it involves the segmentation of prostate volumes through the use of a custom made 3D network based on a CNN model, specifically the UNet. The prostatic volumes are initially preprocessed to remove the bias field, a noise specific for the MRI acquisition modality, and to normalize the intensities. The 3D network is used as it allows for greater spatial coherence, as opposed to 2D networks which are trained on single slices, losing the information deriving from the slices' connections within the volume. The network's outputs will be subsequently refined using an Active Shape Model, that is a statistical model that uses parameters calculated on a set of training images to modify the contours of the segmentation to render them closer to the real contours of the prostate.\\ This model is trained on manual volumetric labels to define a Mean Shape Model that will guarantee the plausibility of the resulting shapes during the testing phase and a series of parameters related to the gray levels around the prostate borders that will allow to iteratively modify the position of the surface points of the CNN output in space to try to improve the segmentation. A comparison with a 2D CNN model will be proposed to evaluate the differences and strengths of a model rather than the other.