Deep learning for medical image analysis

Artificial intelligence is a sector characterized by the development of algorithms through which it is possible to analyze complex data, mimicking the human process of learning and applying the acquired information. A sub-category of artificial intelligence is the deep learning. By freeing up the analysis process from pre-processing techniques based on a priori knowledge of the human operator, deep learning allows the automatic extraction of salient information from "raw data" by using a series of levels of representation. This makes it a sector of particular interest for the research world, as consequence of the possibility to identify new information characterizing the data, but at the same time also for the industry world thanks to the ability of analyzing large amounts of data increasingly available and accessible. In few years these methods have revolutionized the computer vision field, thanks also to a class of algorithms that has shown remarkable performance in the analysis of the images, i.e. the convolutional neural networks. The use of these tools has been widely used, not only in the field of natural image analysis, but also in the field of medical imaging for the extraction and quantification of clinical information. In this context, one of the most critical operations, that is carried out within the analysis of medical images, is represented by the segmentation, i.e. the process that leads to semantically separate different structures within an image. This operation can be exploited by us- ing different methods, as there is not a universal algorithm to perform segmentation in medical images. The choice of the best strategy will depend on the image modality, the anatomical region that is going to be considered and from what is defined the clinical task. Given the high variability of competing factors for the choice of the algorithm to be used, in clinical practice manual segmentation strategies or semi-automatic approaches are so used, which however are resolved in time consuming processes and often prone to errors, due to the high dependence on the operator and his degree of experience.
Convolutional neural networks allow in this sense to overcome these limits, providing a general and robust method to perform multimodal segmentation operations. In this doctoral thesis, a series of approaches have been developed, based on the use of convolutional neural networks, for the analysis of medical images. In particular, three different types of acquisition were treated, such as computed tomography (CT) acquisitions, with and without contrast medium, magnetic resonance images and mammograms. The cardiac district is considered the area of interest for both CT and magnetic resonance. The CT scans were acquired for the visualization of anatomical and structural information, whilst the magnetic resonance images for functional information studies. All the proposed methods have been designed with the aim of solving a specific medical task, with the development of systems able to extract and quantify automatically clinical information, providing an accurate and fast alternative to manual or semi-automatic solutions. The main contributions of this work on CT images are: the development of a system for the analysis of coronary calcium, a system for the quantification of pericardial fat and a method for the synthesis of data with contrast medium starting from an image that is deprived of it. Regarding magnetic resonance acquisitions, a system for the automatic identification of the left ventricular wall and able to divide it into six segments has been proposed. Finally, a model for the identification of microcalcification clusters was developed for mammographic images. The obtained results have highlighted a high system capability to imitate the expert radiologist work, providing general prediction on new coming data, obtaining reliable measurements and allowing their use for large-scale studies.