Artificial intelligence-based organ and tumour segmentation in prostate cancer patients. Studies on PET/CT and pre-treatment CT scans.

Abstract

Aim: This thesis aimed to evaluate the prognostic and clinical significance of artificial intelligence (AI)-based segmentation and quantification of the prostate, the prostate tumour and the surrounding risk organs (OARs), i.e., urinary bladder and rectum on positron-emission tomography/computed tomography (PET/CT) images, and pre-treatment CT scans of prostate cancer patients (PCa). Material and methods: In Paper I, 100 CT scans were used as a training set, regarding prostate segmentation. A separate set of 45 PET/CT scans of PCa patients was used as a test set, regarding both the volume of the prostate and the prostate tumour. AI-based and manual measurements were compared using the Dice similarity coefficient (DSC) and Bland-Altman analysis. In Paper II, all scans from Paper I were used as a training set, while a separate cohort of 304 PET/CT scans of PCa patients comprised the test set, for AI-based quantification of the prostate gland and the tumour, which were associated with survival using Cox proportional hazards regression and survival analysis (Papers I and II). Another AI model was trained to segment the prostate and OARs using 1070 pre-treatment CT scans, while its performance was tested on additional 348 scans. DSC, Hausdorff distance (HD) and mean surface distance (MSD) were performed to compare the AI-based segmentations with the corresponding manual delineations (Paper III). In Paper IV, 120 AI-based organ segmentations from the test set of Paper III, along with their corresponding manual delineations were evaluated by three radiation oncologists, regarding their clinical applicability, using Visual grading characteristics analysis. Results: The agreement between AI-based and manual prostate volume measurements was good with DSC > 0.7 for most cases (Paper I). The AI-based PET/CT measurements were significantly associated with patient survival (Papers I and II). Most of the AI-based organ segmentations showed good agreement with the corresponding manual delineations (DSC > 0.7, HD 3-162 mm, MSD 0.5-8 mm) (Paper III) and were assessed as needing no or minor changes to be clinically acceptable (Paper IV). Conclusion: AI models could offer significant potential for improving patient stratification and treatment planning of PCa patients. However, human intervention remains essential to reach clinical acceptability with the AI model in Papers III and IV.