@phdthesis{1fcab5cb713340e59a8cb9c212b57b61,
title = "Artificial Intelligence-based Assessment of Prostate Cancer Metastases in PET/CT",
abstract = "Background: Quantification of tumor burden from bone scan in the form of automated Bone Scan Index (aBSI) has been validated as an imaging biomarker for patients with prostate cancer. Positron emission tomography combined with computed tomography (PET/CT) is more sensitive and accurate compared to conventional imaging such as bone scan. The evaluation of medical images including PET/CT is challenged by time-consuming and subjective analysis, and issues with intra- and inter-reader agreement. An objective and quantitative method for analysis of PET/CT, similar to aBSI for bone scan, is an unmet need. Artificial intelligence (AI) has the potential to meet this need.Aim: The aim of this thesis was to develop AI-based models for automated quantification of skeletal tumor burden, and for detection of pelvic lymph node lesions, in PET/CT scans from patients with prostate cancer.Methods: In paper I, standardized uptake value (SUV)-based PET indices reflecting the whole-body tumor burden in [18F]fluoride PET/CT scan were calculated in a group of patients with prostate cancer, and compared to aBSI. For proof of concept, the association of PET index with overall survival was investigated. In paper II, convolutional neural networks (CNNs) were trained to segment the axial skeleton in CT scans. In paper III, CNNs were trained to segment bone lesions in [18F]fluoride PET for automated quantification of the skeletal tumor burden. The PET indices in paper I and III were also compared to a SUV threshold for lesion segmentation. In paper IV, a CNN was trained and tested to detect pelvic lymph node lesions in [18F]PSMA-1007 PET/CT scans. The performance of all models was evaluated against readers used as reference.Results: Paper I showed that the SUV-based PET index was associated with overall survival. In paper II, the AI model for bone segmentation performed equally well as an experienced reader. Paper III showed that the performance of an AI-based model for segmentation of bone lesions and quantification of skeletal tumor burden was more similar to experiened readers than a global SUV threshold. The sensitivity for the AI model in paper IV was in level with that of experienced physicians. Inter-reader disagreement among experenced physicians was seen in paper III-IV.Conclusions: AI-based models for automated assessment of PET/CT were developed. Tumor burden measured from PET/CT carries prognostic information in patients with prostate cancer. The difficulty in achieving high inter-reader agreement emphasizes the need for automated and objective scan interpretation. AI-assisted quantification of tumor burden holds potential as a future prognostic imaging biomarker for patients with prostate cancer. Increasing the amount and variety of training data should further enhance the performance of the proposed models.",
keywords = "Artificial intelligence (AI), Convolutional neural networks (CNN), PET/CT, Prostate cancer, Automated, Tumor burden, Bone metastases, Lymph node metastases",
author = "{Lindgren Belal}, Sarah",
note = "Defence details Date: 2022-12-09 Time: 13:00 Place: Rum 2005/2007, Carl Bertil Laurells gata 9, Sk{\aa}nes Universitetssjukhus i Malm{\"o}. Join by Zoom: https://lu-se.zoom.us/j/62799851380 External reviewer Name: Riklund, Katrine Title: Professor Affiliation: Department of Radiation Sciences, Ume{\aa} University, Sweden",
year = "2022",
language = "English",
isbn = "978-91-8021-325-7",
series = "Lund University, Faculty of Medicine Doctoral Dissertation Series",
publisher = "Lund University, Faculty of Medicine",
number = "2022:163",
type = "Doctoral Thesis (compilation)",
school = "Department of Translational Medicine",
}