Diffuse large B-cell lymphoma (DLBCL) is an aggressive cancer worldwide. Since they can be metastasized, whole-body FDG PET/CT is particularly useful for evaluating the burdens of metabolically active lesions. These measurements require accurate lesion segmentation on the PET/CT. Deep-learning (DL) approach has shown a great potential for automating
the lesion segmentation, however its performance may decrease on external datasets. In this study, we develop and assess a framework for accurate quantification of metabolically active
DLBCL lesions. The framework includes a proposed algorithm, Click-to-Grow, which enables human-in-the-loop correction of under-segmented lesions by the DL segmentation output via semi-
automated threshold-based methods without retraining the DL model. The experimental results on 28 PET/CT images from a local dataset suggested that AI combined with 41% SUVmax
refinement strategy obtained a mean DSC of refined lesion segmentation of 95%. In addition, the obtained medians of TLG and total-body TLG errors (tTLG) of the strategy are 6.61
and 18.63 g/mL × cm³, respectively. Moreover, Wilcoxon signed-rank tests suggested that the 41% SUVmax refinement obtained statistically significantly smaller TLG and total-body TLG errors
than those from the AI only segmentation outputs (medians of 20.35 and 30.56 g/mL × cm³, respectively; p-values ≤ 0.01). In conclusion, via improving the lesion segmentation accuracy,
the threshold-based refinement algorithm is able to improve the DLBCL lesion burden quantification accuracy, enabling potential use of the AI in the clinical pratise.

PET/CT segmentation,diffuse large B-cell lym- phoma,region growing,total metabolic tumor volume,interactive correction