Deep Learning based extension for ultrasound imagery

Medical imagery has been on a continuous development in the last decades, as computers have gotten more powerful, with more potent software. Nowadays, medical staff can perform 3D reconstruction of various anatomical structures, ultrasound image representation with various devices such as the CT scan1[1] or ultrasound image processors. The ultrasound representation of an image allows the examiner to see inside a person’s body, which facilitates the detection of malicious cells or tumors. However, even though medical imagery is on a steep development curve, it still lacks usability due to poor image quality. Because of this, doctors consider it a challenge sometimes to detect various anatomical structures that can be vital towards a patient’s treatment. In recent years, deep learning has been a popular choice for automatizing processes. From major corporations wanting to automate their business processes to simple real world solutions, deep learning can empower every kind of implementations, providing new dimensions to every project. With the development of Computer Vision techniques, software engineering has seen the brith and rise of Segmentation algorithms [2]. This thesis aims to provide an assistive system for Ultrasound Imagery, through the means of Deep Learning based Segmentation and general Computer Vision based techniques. Software Engineering is a relatively fertile ground in terms of autonomous products which are used by developers or team managers. Essentially, it contains processes and techniques which are frequently used in development. Since Deep Learning is more than data analysis and interpretation, as it is with every type of development process, Software Engineering can make the construction of Deep Learning applications more efficient, with more user focus and an overall better user experience. This work aims to prove that the principles used in Software Engineering can enhance Deep Learning products, through the means of a medical focused application. The solution make use of a Semantic Segmentation architecture that can accurately2 segment and classify tumor-like structures in ultrasound images of the breast area. At the same time, the system is able to perform visual effects on the image produced, such as brightness and contrast adjustments, in order to boost usability. The system has been validated with real-world users, by various techniques frequently seen in the Software Engineering area