Artificial Intelligence methods for Physics applied to Medicine

The aim of the present PhD project is the development and validation of Artificial Intelligence(AI) models for applications in the medical field, specifically within the domain of MedicalPhysics. Three main applications are addressed: (i) the development and validation of asegmentation model for the Ventral Intermediate Nucleus (VIM), designed to support EssentialTremor treatments, enable real-time treatment planning, and enhance personalized medicine;(ii) the development of three Quality Assurance (QA) protocols, along with a dedicated softwarepipeline, for AI-powered auto-contouring tools in Radiotherapy; and (iii) the development ofPhysics-Informed models capable of mitigating the blurring effect in dosimetry applicationsinvolving Fricke hydrogels.Essential Tremor (ET), the most common movement disorder in adults, significantlyimpacts daily activities due to involuntary tremors affecting the hands, head, voice, or limbs,thereby reducing quality of life. Although its exact pathogenesis remains unclear, when pharmacologicaltreatments fail, surgical interventions targeting the VIM nucleus of the thalamus caneffectively suppress tremor. Several treatment techniques are available, varying in invasiveness;regardless of the technique adopted, treatment planning based on Magnetic Resonance imagingis mandatory. The objective is to identify a Region of Interest (ROI) encompassing the VIM.Existing methods for VIM identification are primarily atlas-based or tractography-based, eachwith specific advantages and limitations. The aim of application (i) was to develop, train, andvalidate a tractography-informed Deep Learning model capable of identifying the VIM in a shorttime while preserving patient specificity.The second application concerns Radiotherapy treatment planning, where the absorbed doseof ionizing radiation must be accurately evaluated prior to treatment delivery. A comprehensiveevaluation of treatment efficacy and associated risks requires accurate contouring of both tumorvolumes and organs at risk (OARs), a task that is typically time-consuming as it is performedmanually by clinicians. Following the Deep Learning revolution, several AI-powered tools havebeen introduced to assist in this process. In this framework, the author’s main contributionin application (ii) was the design and development of three QA protocols for a clinicallyimplemented auto-contouring software.Fricke hydrogels belong to the class of chemical dosimeters and are of great interest inRadiotherapy dosimetry due to their ability to map spatial dose distributions. Despite theiradvantages, they are affected by ion diffusion, which leads to a blurring effect that limits theirtemporal applicability. Most existing research has focused on optimizing chemical compositions toreduce diffusion. In contrast, the work presented in application (iii) explores a post-measurementAI-based approach to mitigate this effect in the framework of the Physics-Informed MachineLearning.Overall, the results obtained during this PhD program support the clinical translation ofAI in the medical field, while emphasizing the necessity of rigorous validation procedures andthe continuous human oversight required to ensure model reliability and safety.