Deep learning-based automated segmentation of 18F-FDG PET images in Hodgkin Lymphoma: A multi-center validation study using clinical trial data form HD16, HD17 and H18

Ziel/Aim: Therapeutic stratification in Hodgkin disease (HD) has traditionally relied on manual assessment, including the Ann Arbor classification and additional risk factors. Quantitative parameters derived from 18F-FDG PET scans like metabolic tumor volume (MTV) and maximum disease extent (Dmax) offer potential continuous variables for refined risk assessment. However, quantification of those metrics is too time-consuming for routine clinical practice. Therefore, the aim of this work is to evaluate the performance of different Deep learning-based segmentation strategies in clinical trial data.

Methodik/Methods: All patients with available baseline 18F-FDG PET from the HD16 (early favorable), HD17 (early unfavorable) and HD18 (advanced) German Hodgkin Study Group trials were included. These multicenter studies recruited patients at 134 sites between August 2008 and December 2016. Lymphoma masses were manually segmented with a threshold of SUV>4. These masks were used to train a deep learning-based segmentation model. The model validation used PET data from medical centers separate from those that provided the training data.

Ergebnisse/Results: The deep learning-based segmentation model was trained on 458 patients, 91 patients were used as external test set (total n=549). The lesion segmentation sensitivity was 87.6% in the total test set (Dice score 87.2%). Lesion segmentation sensitivity was 86.8% in the HD16 subset, 97.6% in the HD17 subset and 83.7% in the HD18 subset. Manually and automatically determined tumor volumes showed strong correlation (HD16: rho=86.9%; HD17: rho=99.5%; HD18: rho=90.3%; for all: p<0.001). Likewise, the Dmax parameter showed strong correlations (HD16: rho=92.9%; HD17: rho=85.8%; HD18: rho=78.6%; for all: p<0.001).

Schlussfolgerungen/Conclusions: The presented deep learning-based approach shows robust performance in automated segmentation of lymphoma manifestations on 18F-FDG PET at different stages of HL, achieving high accuracy in both volume quantification and dissemination measurement. The strong correlations between automated and manual measurements for MTV and automated PET analysis can potentially facilitate the routine clinical use. Future studies on the prognostic value of the automatically derived quantitative parameters are warranted.

Publikationsverlauf

Artikel online veröffentlicht:
12. März 2025