Abstract
Objective:To evaluate MedNeXt, a deep learning network, for automated segmentation of 18F-FDG PET images in angioimmunoblastic T-cell lymphoma (AITL) and validate the prognostic utility of predicted total metabolic tumor volume (pTMTV).
Methods:A multicenter cohort of 84 AITL patients was stratified into training (n=68) and validation (n=16) sets (8:2 ratio). MedNeXt was trained using a Dice-cross-entropy dual-loss function. Segmentation accuracy was quantified through seven metrics: Dice similarity coefficient (DSC), Jaccard index (JSC), sensitivity (SEN), positive predictive value (PPV), false discovery rate (FDR), 95% Hausdorff distance (HD95), and average symmetric surface distance (ASSD). Prognostic significance of pTMTV was assessed via Cox proportional hazards models.
Results:Validation cohort performance metrics (mean±SD): DSC=0.769±0.153, JSC=0.645±0.173, PPV=0.773±0.201, FDR=0.227±0.201, SEN=0.799±0.110, ASSD=1.903±2.288 mm, HD95=32.183±44.664 mm. Ground truth TMTV (gtTMTV=949.4±734.5 cm³) and pTMTV (934.1±617.7 cm³) demonstrated strong agreement in Bland-Altman analysis and linear regression (R²=0.868, p<0.001). Elevated pTMTV (≥606.9 cm³) independently predicted reduced progression-free survival (HR=1.781, 95% CI:1.057-3.002; p=0.030) and overall survival (HR=2.190, 95% CI:1.042-4.600; p=0.038).
Conclusion:MedNeXt achieves precise lesion delineation and automated quantification of TMTV in AITL PET imaging. pTMTV emerges as an independent prognostic predictor for survival outcomes, demonstrating clinical viability for lymphoma management.
