AI-based medical image analysis and interpretation: from feature extraction to decision support

Over the past few years, we have witnessed an explosion of highly performing Artificial Intelligence (AI) models addressing diverse tasks in computer vision within the healthcare domain. However, their integration into everyday clinical practice remains limited. The field of AI-based medical image analysis faces multiple challenges, including a scarcity of data, variable image quality, and the imperative for interpretable and generalizable results. Conversely, the potential benefits of employing such technology in routine clinical practice are extensive. These include the possibility of seamlessly incorporating fully automated decision support systems at different stages of the clinical routine, ranging from early diagnosis to prognosis prediction. This thesis aims to delineate a comprehensive workflow for building fully automated and easy to customize AI-based medical image analysis pipelines. Three distinct case studies, designed and analyzed in collaboration with highly specialized European centers, are presented. Each case pertains to a specific medical domain - oncological, cardiological, or neurological - presenting unique challenges from both clinical and technical perspectives. The proposed pipelines are crafted to meet specific criteria: high performance, reproducibility, ease of generalization, and interpretability by the final clinical user, who must view the system as trustworthy, even without expertise in the technical implementation. Additionally, the applications have been meticulously designed to demand limited computational resources while maintaining optimal performance. The first two case studies present fully automated systems designed to enhance the efficiency and diagnostic accuracy during screening programs. The first system accurately identifies malignant microcalcifications from mammograms during breast screening programs to mitigate the high false positive rate. In the second case, a quick and accurate automated system is introduced to rule out patients requiring further clinical investigations during coronary artery disease screenings, based on the degree of occlusion of the three main coronary arteries visible from cardiac CT angiography. These pipelines are specifically crafted to alleviate time-consuming and operator-dependent tasks. In the last case study, an easily generalizable workflow for prognosis prediction and biomarkers discovery is discussed. The presented pipeline is capable of identifying novel imaging biomarkers from follow-up MRI sequences with the objective of predicting poor long-term functional outcomes in acute ischemic stroke patients. This kind of system has the potential to fully exploit the information content in routinely acquired clinical images, providing insights into the pathophysiological mechanisms of the disease and predicting its possible evolution. This goes beyond qualitative biomarkers or simple lesion measurements, which are often the only indicators used to guide the best clinical intervention. Finally, the recent emergence of generalist foundation models that are gradually shifting the landscape of AI is deeply discussed. In the medical domain, this approach holds significant promise in overcoming common limitations, particularly those related to data quantity and quality. An evaluation study designed to test the adaptation of a general-purpose segmentation algorithm on a real dataset of patients with hemorrhagic ictus is presented. The high performance of the implemented system showcases a novel and easy-to-implement approach for expediting the manual hematoma delineation process from CT scans acquired in emergency rooms. If, on the one hand, this demonstrates the great potential of the generalist approach, it is also undeniable that various concerns, particularly from legal and ethical perspectives, must be promptly addressed to ensure the safety of the final supporting tools improving healthcare accessibility, fairness, and inclusivity.

Negli ultimi anni, abbiamo assistito a un'enorme diffusione di modelli di Intelligenza Artificiale (IA) ad elevate prestazioni che affrontano diverse sfide nel campo della visione artificiale in ambito biomedico. Tuttavia, l'integrazione clinica di queste tecnologie è ancora limitata a causa di sfide come la scarsità di dati e la necessità di risultati interpretabili. La tesi propone la creazione di pipeline automatizzate per l'analisi di immagini cliniche utilizzando l'IA, con tre casi studio in ambito oncologico, cardiologico e neurologico. Le pipeline mirano a performance elevate, garantendo allo stesso tempo riproducibilità, interpretabilità e facilità di generalizzazione. I primi due casi riguardano sistemi per migliorare l'efficienza diagnostica nello screening di microcalcificazioni mammarie maligne e malattie coronariche, rispettivamente. Nel terzo caso, viene presentato un workflow per la predizione della prognosi e l’identificazione di nuovi biomarcatori utilizzando sequenze di risonanza magnetica di pazienti con ictus ischemico acuto. Infine, viene approfondito il tema dell'impiego di modelli generalisti o “fondativi" che sta gradualmente cambiando il panorama dell'IA. Nel dominio medico, questo approccio promette di superare limitazioni comuni, in particolare quelle legate alla quantità e qualità dei dati. Viene presentato uno studio di validazione disegnato per testare l'adattamento di un algoritmo di segmentazione progettato per un uso generale su un set di dati reale di pazienti con ictus emorragico. Le alte prestazioni ottenute da tale sistema mostrano come questo nuovo approccio sia facile da implementare e possa quindi accelerare il processo di segmentazione manuale dell'ematoma dalla TAC di pronto soccorso. Se, da un lato, questo dimostra il grande potenziale dell'approccio generalista, è anche innegabile che diverse sfide, in particolare legate alla sfera medico-legale ed etica, devono essere affrontate tempestivamente per garantire la sicurezza di tali software al fine di arrivare a migliorare l'accessibilità, l'equità e l'inclusività nell'assistenza sanitaria.