Artificial Intelligence for 18F-FDG PET/CT prognosis in diffuse large B-cell lymphoma

DLBCL is the most common subtype of non-Hodgkin lymphoma and remains a major clinical challenge due to its biological and clinical heterogeneity. Despite major therapeutic advances with rituximab-based immunochemotherapy, approximately one third of patients relapse or experience early progression, highlighting the need for improved risk stratification beyond established clinical scores such as the International Prognostic Index (IPI). Baseline 18F-FDG PET/CT is central to staging and response assessment and provides quantitative biomarkers, including total metabolic tumor volume (TMTV), SUV-derived metrics, and dissemination measures, that correlate with outcome. However, these biomarkers are highly dependent on tumor segmentation and on technical factors related to PET acquisition and image reconstruction, limiting reproducibility across centers. This thesis investigates and addresses these limitations through two complementary approaches. In Part I, we evaluate the impact of different PET reconstructions (EARL1, EARL2, and high-resolution clinical protocols) on segmentation outcomes and PET-derived biomarkers. We demonstrate that reconstruction settings can substantially alter TMTV estimates, with higher-resolution reconstructions generally yielding smaller volumes. Statistical harmonization using ComBat improves inter-reconstruction agreement but does not fully compensate for spatial and visual differences in the images, underscoring the importance of harmonizing acquisition and reconstruction in line with EANM standards. We further assess automated lesion detection/segmentation using deep learning (LIONZ) and propose a hybrid strategy (LIONZSUV4) combining AI-based detection with a robust fixed-threshold approach (SUV4.0), resulting in more consistent biomarker extraction and reduced observer dependence in multicenter settings. In Part II, we develop and validate end-to-end deep learning models that predict progression risk directly from baseline PET images without manual segmentation. A 2D convolutional neural network (CNN) using maximum intensity projections (MIP-CNN) predicts 2-year time-to-progression and outperforms the IPI. Tumor ablation experiments confirm that predictions are driven by tumor-related signal rather than spurious features. We then quantify the sensitivity of model outputs to reconstruction differences and show that image-based harmonization (e.g., resolution matching) stabilizes predictions more effectively than purely statistical correction. External validation across multiple international clinical trial cohorts demonstrates robust generalization (AUC ~0.60–0.66) and performance comparable to advanced radiomics models, while remaining fully automated. Finally, we extend this approach to 3D CNNs trained on full PET volumes and incorporate explainable AI (3D occlusion maps) to localize image regions contributing to predictions, improving interpretability for clinical translation. Overall, this work shows that robust prognostic modeling in DLBCL using PET/CT requires careful standardization and harmonization of imaging, and that deep learning can provide accurate, scalable, and interpretable risk prediction that reduces reliance on manual segmentation and supports future precision-medicine workflows.