(eXplainable) Artificial Intelligence in Positron Emission Tomography imaging

This thesis explores various applications of (X)AI in PET imaging across different medical conditions and radiotracers. The AI applications discussed range from pre‑processing tools designed to reduce acquisition time, lower injected doses, enhance image quality, and improve image alignment, to diagnostic and prognostic tools aimed at detecting pathologies and predicting clinical outcomes. The first part of this thesis investigated AI as a pre‑processing tool for 18F‑FDG and 89Zr‑antibody PET images in patients with different cancer types. In Chapter 2, we aimed to reduce acquisition time and injected activity of 18F‑FDG PET while maintaining diagnostic value. For this purpose, low‑count 18F‑FDG PET images with low SNR were acquired. Since 89Zr‑antibody PET is also characterized by low SNR, we subsequently aimed to improve SNR—and thus diagnostic value—for both low‑count 18F‑FDG and full‑count 89Zr‑antibody PET images. We adopted an AI model and compared its performance with conventional denoising methods. The AI model improved SNR in both modalities, achieving diagnostic performance comparable to or better than the original PET images, and outperforming conventional denoising approaches. In Chapter 3, we evaluated the impact of an AI‑based PET‑CT co‑registration model on 18F‑FDG PET quantification using ldCT‑based VOIs in patients with MM. Traditional intensity‑based co‑registration methods struggle to align PET and ldCT images due to differences in image characteristics. To address this, we used AI to generate an sCT image to improve inter‑voxel correspondence with ldCT. Using the sCT, we demonstrated improved alignment between 18F‑FDG PET and ldCT, as well as significant effects on PET quantification using ldCT‑based VOIs. Both studies demonstrated proof of concept, but further validation is required across additional radiotracers, PET/CT systems, clinical centres, and patient cohorts. The second part of this thesis describes the development of (X)AI‑based prediction models. In Chapters 4 and 5, we examined current XAI methods in radiology and nuclear medicine and, based on this assessment, proposed a direction for a modality‑specific XAI model. We concluded that no clear consensus exists on how XAI should be evaluated to bridge the gap between medical professionals and AI for clinical use. We advocated for systematic technical and clinical quality validation of XAI methods. Furthermore, to ensure unbiased and safe integration of XAI into clinical workflows, data minimization (including anatomical data) and quality control procedures should be incorporated. In the subsequent chapter, we developed a PET‑CT–based XAI model aligned with these principles, aiming to provide technically and clinically sound potential imaging biomarkers. As proof of concept, we used pre‑treatment 18F‑FDG PET to predict dichotomous PFS in patients with mCRC. Extensive quantitative and qualitative assessment showed that the XAI model performed well in predicting PFS and that both tumour and non‑tumour variables appeared to be associated with PFS. Although we did not demonstrate that these variables improve prognostication in this cohort, the study offers a potential direction for systematic technical and clinical evaluation of XAI methods in PET imaging. Finally, Chapter 6 presents an AI model to predict positive or negative 18F‑Florbetapir brain PET studies in patients with SCD. Visual reading of these scans is part of the diagnostic process but can be time‑consuming and susceptible to inter‑ and intra‑observer variability, especially in borderline amyloid pathology. We aimed to develop an AI model replicating this diagnostic process. The model achieved high performance in predicting amyloid status, but unlike Chapter 5, no systematic technical or clinical assessment was performed. Therefore, despite promising results, similar rigorous evaluation is required to further validate the model.