Deep Learning based Detection and Tracking of the IVC in Ultrasound Scans

This work of thesis originated within a start-up company, VIPER s.r.l., which was established as a research project of the Polytechnic of Turin, which developed a semi-automatic software to delineate and trace the displacements of the edges of the Inferior Vena Cava (IVC) not in a single direction, but in both transverse and longitudinal sections in ultrasound videos, meaning in a totally non-invasive way. This software allows the computation of different measurements, such as diameters, areas, Caval Index (CI) and so on, based on the assessment of the pulsatility of the IVC itself. These measurements can lead to the evaluation of the volemic state of a patient and the pressure present in the right atrium (Right Atrial Pressure, RAP), which are two different indicators that can be valuable in many fields of medicine, ranging from internal medicine, in order to distinct between hypovolaemic and euvolemic patients, to cardiology, for the evaluation and management of heart failure, and many more. One limitation of this software is the impossibility of obtaining these measurements in real-time: in fact, at first the ultrasound video is obtained and only in post-processing the video can be elaborated, and all the measurements of interest therefore obtained. Furthermore, as previously said, it is a semi-automatic software, meaning it needs the intervention of an operator to start working. This thesis arose out of the need to make this software work in real-time and to make it completely operator-independent. To make it possible, it was studied a deep learning methodology that allows the automatic detection of the IVC in real-time. In particular, it was chosen to use a YOLOv8 model, which is the current state-of-the-art of the real-time object detection models: the aim of this tool is to identify and classify Inferior Vena Cava within ultrasound videos, both in transverse and longitudinal sections with high accuracy. Furthermore, in order to improve the ability of the object detection model to follow the IVC movements throughout the ultrasound video, it was chosen to couple it with an object tracking model, in particular it was opted for DeepSORT, a very popular object tracking algorithm, also based on deep learning. Therefore, the discussion will begin with an initial descriptive part that will cover anatomy and physiology of the IVC, ultrasounds basics, and an overview about object detection and the most famous algorithms used nowadays. In the second part, they will be discussed the methods that were used in this work of thesis, in particular the ones concerning YOLOv8 and DeepSORT will be explored in detail. Finally, they will be exposed the results achieved in this work of thesis.