Evaluation of an artificial intelligence method for lesion segmentation of baseline FDG PET studies of DLBCL patients

Background. The aim of this study is to investigate the use of an artificial intelligence (AI) method, LIONZ, in combination with an intensity-based threshold method, SUV4.0, for the automatic selection and segmentation of diffuse large B cell lymphoma (DLBCL) lymphoma lesions. Methods. 296 DLBCL ¹⁸ F-FDG PET scans were analyzed. Metabolic tumor volume, peak standardized uptake value (SUVpeak) and, maximum distance from the bulkiest lesion to another lesion (Dmaxbulk) were extracted from the LIONZ and LIONZ ^SUV4 segmentations and compared to those extracted from SUV4.0 segmentations using Pearson correlation (p < 0.05) and Bland-Altman plots. Segmentation performance was assessed using the Dice similarity coefficient (DSC) with SUV4.0 segmentation as a reference. A prediction model which includes MTV, SUVpeak, Dmaxbulk, age and performance status was used to predict the probability of 2 year time to progression using the parameters extracted from the LIONZ, LIONZ ^SUV4 and SUV4.0 segmentations. Association of probabilities was evaluated using Pearson correlation (p < 0.05) and Bland-Altman. The area under (AUC) the curve was used to assess and compare the performance of both methods. Results. The median DSC (interquartile range) for LIONZ when compared to SUV4.0 was of 0.77 (0.64–0.84) and for LIONZ ^SUV4 of 0.87 (0.80–0.93). MTV, SUVpeak and Dmaxbulk from both the LIONZ and LIONZ ^SUV4 were highly correlated to the SUV4.0 segmentations derived parameters (R ≥ 0.80, p < 0.0001). LIONZ ^SUV4 reduced overestimation of segmented areas and LIONZ ^SUV4 MTV showed a stronger agreement with that of SUV4.0 compared to LIONZ (0.99 and 0.80 respectively, p < 0.0001). The prediction model yielded an AUC of 0.74, 0.78 and 0.79 when using segmentations from LIONZ, LIONZ ^SUV4 and SUV4.0 respectively. The predicted probabilities yielded by the models using the LIONZ and LIONZ ^SUV4 segmentations were also highly correlated with those of SUV4.0 segmentation (0.9 and 0.96 respectively, p < 0.0001). Conclusion. LIONZ ^SUV4 segmentations highly overlapped with those of SUV4.0. LIONZ ^SUV4 led to a stronger agreement of PET parameters and predictions with SUV4.0 compared to LIONZ. Overall, LIONZ ^SUV4 is a suitable method for DLBCL lesion segmentation and potentially decreases reader-variability compared to threshold only based segmentation methods.