Computer Assisted Diagnosis in PET/CT

Abstract

Artificial intelligence (AI) has revolutionised problem solving in a wide range of industrial as well as research domains. Particularly, computer-aided diagnosis (CAD) and clinical decision support systems (CDSSs) as sub-domains of AI, have gained critical importance in many biomedical and clinical domains such as virology, computational neuroscience, and oncology. As making accurate decisions in a timely manner is an inevitable part of daily routines in the medical and clinical domains, machine learning (ML) and deep learning methods are widely applied in CAD and CDSSs to provide diagnostic and prognostic assistance for the researchers and physicians as the domain experts. <br /> Focusing on advanced prostate cancer (PCa) disease as an example, the procedure of disease staging and patient screening using established CAD tools is considered time consuming and attention intensive. In many clinical practices, this procedure includes examining patients’ prostate-specific membrane antigen-positron emission tomography/computed tomography (PSMA-PET/CT) scans and analyzing patient-specific clinical factors in a daily routine. Thus, as the main motivation behind this PhD thesis project, AI and ML based methods are utilized to automate the corresponding diagnostic and prognostic pipelines. <br /> Accordingly, providing an automated CDSS which facilitates: 1) visualization and annotation of medical scans, 2) automated segmentation of pathological uptake, 3) prediction of treatment outcome taking advantage of radiomics features extracted from Gallium[68]-(68Ga)-PSMA-PET/CT scans in PCa patients was the main objective of this thesis. <br /> To this end, we introduce AutoPyPetCt, an automated pipeline developed in Python which takes multimodal whole-body baseline 68Ga-PSMA-PET/CT scans and patient-specific clinical parameters as input and applies state-of-the-art statistical, ML, and deep learning techniques to automatically identify and segment pathological uptake all over the body, to anticipate responders to Lutetium[177]-(177Lu)-PSMA therapy, and to predict overall survival of the PCa patients. <br /> To achieve this, on the one hand, multimodal PET/CT scans integrate functional as well as anatomical aids to locate malignancies as volumes and regions of interest (VoIs and and RoIs respectively). On the other hand, a variety of conventional parameters (such as standardized uptake value (SUV)) as well as radiomics features (such as textural heterogeneity features) extracted for the VoIs/RoIs together with patient-specific clinical factors (such as age and prostate-specific antigen (PSA) level) form the basis for statistical and ML-based analyses towards prognostic hypotheses realizing the prediction of patient level outcomes such as treatment response and overall survival. <br /> The main contribution of the methods is to provide automated decision support tools to manage patients with advanced PCa in shorter times and with limited annotation effort. To investigate the relevance and to quantify the performance of the methods, multiple retrospective quantitative as well as qualitative clinical studies have been conducted which resulted in several preliminary conference abstracts, four journal papers, and one conference paper. The studies had been carried out along the whole project’s life-cycle, starting by a proof of concept and finalizing with the evaluations of the integrated solution pipeline. <br /> The findings from the clinical studies confirmed the overall relevance of the methods and their potential to replace parts of current clinical routine procedures in the future. Most interestingly, the provided automated segmentation tools achieved high performance in true delineation of pathological uptake which outperformed a standard established thresholding based approach. However, the results of the treatment response prediction studies, regardless of different segmentation methods, identified rooms for improvement. <br /> To conclude, the provided automated decision support system has shown its potential to serve as an assistant for the management of patients diagnosed with advanced prostate cancer disease. However, to further assess the generalizability of the findings and to improve the decision making certainty, studies including multicentric data should be considered as future work.