Clinically reliable and stable automated segmentation of DLBCL lesions on PET/CT using self-configuring nnU-Net for...

Accurate lesion segmentation on 18F-FDG PET/CT is essential for the effective management of Diffuse Large B-cell Lymphoma (DLBCL). While deep learning architectures, particularly U-Net and its variants, are widely used, identifying the most reliable and robust framework for clinical quantification remains a challenge. This study aims to perform a systematic, three-way comparative evaluation of representative deep learning approaches to establish a performance hierarchy and clarify the source of superiority.

A systematic comparison was performed among three architectures: a manually configured classic U-Net, an advanced Attention U-Net (incorporating attention gates), and the self-configuring nnU-Net framework, using 217 annotated whole-body PET/CT scans. The classic and Attention U-Nets were explicitly designed and tuned by experts, whereas nnU-Net automated its entire pipeline. All models were evaluated via 3-fold cross-validation using the Dice Similarity Coefficient (DSC), Intersection over Union (IoU), sensitivity, and specificity. Clinical relevance was assessed by comparing model-derived Total Metabolic Tumor Volume (TMTV) with expert annotations using correlation and Bland-Altman analyses. The final ground-truth segmentation for each patient was generated through a consensus process between the two expert annotators.

The results consistently favored nnU-Net across all evaluated metrics, supporting a robust performance hierarchy among the evaluated models. It achieved a higher mean DSC (0.83 vs. 0.61 vs. 0.71), IoU (0.71 vs. 0.52 vs. 0.62), sensitivity (0.85 vs. 0.56 vs. 0.72), and specificity (0.94 vs. 0.87 vs. 0.92). nnU-Net further showed improved stability, reflected by lower performance variability across cross-validation folds. In TMTV estimation, nnU-Net showed substantially reduced systematic error and much narrower Bland-Altman limits of agreement (-15.8 to +10.5 mL vs. -108.5 to +71.1 mL vs. -97.3 to +67.7). These findings suggest that the observed advantages of nnU-Net are primarily associated with comprehensive pipeline optimization rather than architectural modifications alone.

This comparative analysis demonstrates that the self-configuring nnU-Net framework provides superior and more stable segmentation performance compared with both manually designed U-Net variants, enabling more accurate and reliable DLBCL lesion delineation and TMTV quantification. Its automated pipeline and robust performance support its integration into clinical workflows. Future studies will focus on multi-center external validation, correlation with clinical outcomes (e.g., survival), and model interpretability.