Evaluation of the detection accuracy of a fully automated AI-based lesion segmentation tool in whole-body FDG PET/CT for lymphoma.

Abstract

Background: This study evaluated an AI-based lesion segmentation tool using a hybrid 2D-3D deep learning model based on the nnUNet architecture with a ResNet18 backbone, incorporating post-processing to reduce false positives and segmentation-related quantitative errors.

Methods: Whole-body FDG PET/CT scans from 72 lymphoma patients (mean age 39.2 ± 18.6 years; 33 females, 39 males; 43.1% Hodgkin lymphoma, 56.9% non-Hodgkin lymphoma) acquired on a uMI550 digital PET/CT system were retrospectively analysed. Automated AI-based segmentations were compared with manual delineations by experienced physicians. Performance was assessed using F1-score, sensitivity, and positive predictive value. Quantitative agreement for all PET metrics was evaluated using Pearson correlation, concordance correlation coefficient, intraclass correlation coefficient, regression, and Bland-Altman analysis. McNemar's test assessed detection differences. The correlations between DSC and image quality metrics were examined.

Results: A total of 711 AI and 848 manually segmented lesions were analysed. The AI model achieved an excellent DSC (89.2%), sensitivity (82.9) and PPV (96.5) which indicated balanced detection performance. Strong correlations were observed for all quantitative parameters (r > 0.95, p < 0.001; CCC > 0.94; ICC > 0.94; 95% CI: 0.93-0.99), with minimal bias on Bland-Altman analysis. DSC showed a moderate negative correlation with liver SNR (r = -0.40, p < 0.001) and weak correlations with liver SUV metrics.

Conclusion: The fully automated AI-based segmentation tool demonstrated good-to-excellent lesion detection accuracy and high quantitative concordance with manual segmentation in whole-body FDG PET/CT for lymphoma patients. Continued model refinement and region-specific training are warranted to enhance detection performance.