A dual-validation 3D nnU-Net framework with harmonized preprocessing for robust DLBCL segamentation in PET/CT images

Diffuse large B-cell lymphoma (DLBCL) is an aggressive and common subtype of non-Hodgkin lymphoma. The automatic segmentation of DLBCL tumors from positron emission tomography/computed tomography (PET/CT) images remains a significant challenge due to the complexity and variable appearance of tumors. In this study, we developed and evaluated a 3D nn-UNet model for the automatic segmentation of DLBCL lesions to support treatment planning and monitoring. The model was trained on 18F-FDG PET/CT scans from 217 patients. Performance was assessed using geometric metrics, resulting in a mean Dice Similarity Coefficient (DSC) of 0.85, Intersection over Union (IoU) of 0.75, sensitivity of 88.3 %, specificity of 95.7 %, and accuracy of 97.1 %. To establish clinical validity, the Total Metabolic Tumor Volume (TMTV) was derived from both ground truth and predicted segmentations. Bland-Altman analysis demonstrated strong agreement, and linear regression confirmed a high correlation between the volumes. The key novelty of our work lies in a harmonized preprocessing pipeline and a dual-validation strategy that integrates geometric metrics (DSC, IoU) with volumetric and metabolic assessments (TMTV, Standardized Uptake Value (SUVmax)). The results, supported by box plots illustrating metric distributions, confirm the model's robustness, reliability, and potential for clinical utility in managing DLBCL.