Fully automated determination of metabolic tumor volume in Hodgkin lymphoma

Abstract Purpose 18 F-Fluorodeoxyglucose ( 18 F-FDG) Positron-emission tomography (PET) is the standard of care for staging of Hodgkin lymphoma (HL). While metabolic tumor volume (MTV) derived from PET imaging has demonstrated predictive value, its clinical application remains limited. We therefore developed and validated a machine learning (ML) model for fully automated PET-based segmentation in HL. Methods A total number of 1,558 PET scans was used, of which 544 were baseline scans from the multicenter German Hodgkin Study Group (GHSG) trials HD16, HD17, and HD18, with manually segmented MTVs (MTV Manual ). A residual encoder U-Net deep learning model was trained on HL-specific GHSG data (“HL-Net”). For comparison, a model was trained on a multicancer dataset with 1,014 PETs separately (“Auto-PET”) and in combination with HL-specific scans (“HL/Auto-PET”). An additional HL-specific variant (“HL4-Net”) incorporated a SUV≥4 threshold mask as an auxiliary input. All models were validated using a hold-out cohort from non-overlapping external HD-trial sites. Results The HL-Net had higher Dice score, sensitivity, and positive predictive value (PPV) compared to the Auto-PET model (Dice: 0.865 vs. 0.661, p <0.001; Sensitivity: 1.000 vs. 0.727, p <0.001; PPV: 1.000 vs. 0.600, p <0.001). The HL4-Net further improved the Dice score compared to the HL-Net (0.949, p <0.001). MTV ML from HL4-Net and MTV Manual showed very strong correlation (R 2 =0.981, p <0.001). In predicting early treatment response MTV ML achieved a comparable area under the curve (AUC) to MTV Manual (0.791 vs. 0.799, p =0.649). Conclusion HL4-Net was optimal for segmentation of disease manifestations with high accuracy, stressing the need for disease-specific models. MTV derived from HL4-Net can be used as surrogate for MTV Manual with equivalent predictive accuracy.