Gliocare: The Road to Gadolinium Free Glioma Imaging

Despite continued reliance on contrast‑enhanced T1‑weighted imaging (CET1) for glioma diagnosis, treatment monitoring, and follow‑up, concerns about health risks, environmental impact, and financial costs of repeated gadolinium‑based contrast agent (GBCA) use have stimulated the search for alternatives. Besides, contrast enhancement alone does not fully capture viable tumor tissue, highlighting the need for new imaging biomarkers to improve patient outcomes. Chapter 2 presents the PENGUIN survey, which assessed patient attitudes toward GBCA use and MRI follow‑up. Patients generally trusted current imaging protocols but preferred GBCA‑free imaging if diagnostic accuracy could be maintained. Experiences with MRI were influenced by factors such as sex, age, and tumor grade, supporting the need for more personalized imaging strategies. Patients were also positive about current follow‑up frequency, an important consideration if imaging intervals are reduced. Chapter 3 provides a narrative literature review of GBCA reduction research, categorized into guidelines, advanced imaging, and AI approaches. Most published studies were retrospective. The review suggests that universal or subgroup‑specific GBCA reduction is feasible and that imaging intervals could be tailored clinically. ASL appears comparable to dynamic susceptibility contrast (DSC) perfusion imaging. Chapter 4 evaluates a clinically feasible whole‑brain 3T Amide Proton Transfer-Chemical Exchange Saturation Transfer (APT‑CEST) sequence. The method showed good short‑ and long‑term reproducibility in healthy volunteers and glioma patients, with tumor‑to‑normal tissue contrast exceeding scan‑rescan variability. Tumor regions showed higher MTR asymmetry values than contralateral tissue, motivating the prospective inclusion of 100 patients in GLIOCARE. In Chapter 5, descriptive and radiomic features from APT‑CEST and ASL were used to predict IDH status, 1p/19q status, and tumor grade across international datasets. MTR asymmetry features were consistently associated with molecular subtypes, and tumor cerebral blood flow was increased mainly in glioblastoma. Combining perfusion and APT features slightly improved prediction. Cross‑vendor generalizability of APT was limited, though normalization to contralateral tissue reduced inter‑dataset variability. Chapter 6 describes the development of the IMAGO database, including approximately 1,700 in‑house patients and an open‑access dataset of 500 adult‑type diffuse glioma patients. The database supports AI training and validation and is continuously expanded with clinical, histological, and radiotherapy data. Chapter 7 evaluates non‑contrast tumor segmentation in 100 glioblastoma patients. Four human raters segmented tumors using only non‑contrast sequences, and an AI algorithm generated additional segmentations that were fine‑tuned by a fifth rater. AI‑assisted segmentations achieved significantly better Dice scores, Hausdorff distances, and volume errors than manual segmentations. Raters improved over time, but longer segmentation time did not correlate with higher accuracy. Chapter 8 addresses limitations in prior AI research by benchmarking GBCA‑free segmentation and synthesis algorithms on a large open‑source dataset across three clinical tasks: enhancement detection, prognostic assessment, and therapy response evaluation. Six segmentation and four synthesis algorithms were evaluated. Segmentation methods consistently outperformed synthesis approaches in sensitivity, negative predictive value, accuracy, and prognostic relevance. Enhancement volumes derived from segmentation predicted survival, and adding age, sex, and surgery type improved prognostic performance. However, neither approach is ready for clinical use. Chapter 9 introduces Apollo SmartGAD, a triage tool that classifies patients by need for contrast administration. Although sensitivity and specificity were promising, misclassification occurred, particularly in small or typically enhancing tumors. The tool is best considered supportive rather than a replacement for current protocols. In conclusion, this thesis demonstrates that APT‑CEST, and to a lesser extent ASL, offer promising GBCA‑free biomarkers for glioma characterization. AI methods show potential but require stronger clinical integration and larger validation studies. Future research should address vendor variability in APT‑CEST and assess whether APT-CEST better delineates true tumor extent than contrast imaging. Aligning AI development with clinical workflows may ultimately reduce GBCA use while improving patient care.