In the field of cancer management, understanding the burden and spread of
disease is essential for informed clinical decision-making. This thesis focuses on the role
of machine learning in analyzing [68Ga]Ga-PSMA PET/CT scans for the diagnosis and
assessment of prostate cancer (PC).
The first study describes the development of a fully automated deep learning
framework for identifying and segmenting malignant lesions in whole-body PET scans
using [68Ga]Ga-PSMA-11. In a multicenter study involving 412 patients, this model
demonstrated high accuracy in patient-level classification and lesion detection, yielding
promising results in extracting quantitative PET biomarkers for evaluating treatment
response.
The second study investigates the distribution of PSMA in newly diagnosed
prostate cancer and its correlation with serum PSA levels and biopsy Gleason Score
(GS). Analysis of data from 256 patients revealed significant associations between PET-
based parameters and disease characteristics. Machine learning techniques effectively
predicted GS and PSA values, emphasizing the utility of [68Ga]Ga-PSMA PET/CT in
staging and treatment planning.
The third study identifies baseline clinical parameters and PET-extracted
features that influence PSA response, overall survival, and progression-free survival in
patients undergoing treatment with 177Lu-PSMA. Analysis of 125 patients with
metastatic hormone-resistant prostate cancer showed that several parameters
significantly impacted clinical outcomes. These findings underscore the potential of
integrating imaging biomarkers to enhance predictive assessments and guide
therapeutic decisions.
Overall, this thesis presents methods of image analysis and their applications in
improving the diagnosis, staging, and treatment of prostate cancer, highlighting the role
of machine learning in developing effective and robust solutions for clinical
applications.