Deep learning [18F]-FDG-PET/CT‑based algorithm for tumor burden estimation in metastatic melanoma patients under immunotherapy

Abstract

Background and purpose: Artificial intelligence is increasingly used in radiation oncology, yet its application for tumor burden (TB) estimation remains limited. This study evaluated the performance of a [¹⁸F]-fluorodeoxyglucose positron emission tomography/computerized tomography ([¹⁸F]-FDG-PET/CT)-based deep learning model, PET-Assisted Reporting System ("PARS", Siemens Healthineers), for lesion detection, segmentation, and TB estimation in patients with metastatic melanoma undergoing immunotherapy.

Materials and methods: This retrospective study included 165 stage IV melanoma patients who underwent [¹⁸F]-FDG-PET/CT imaging prior to immunotherapy. Gross tumor volumes were segmented using PARS and compared with manual delineations performed by radiation oncologists. Performance was assessed through lesion detection metrics (precision and recall), individual lesion volume agreement, and overall TB estimation accuracy.

Results: PARS demonstrated an overall recall (sensitivity) of 68.9 %, though with modest precision (46.8 %). Performance was location-dependent, with highest precision observed for lung lesions (74.0 %) and lowest for bone lesions (32.9 %). For lesions detected by both methods, PARS tended to underestimate lesion volumes by an average (median) of 0.9 cc (median relative percentage difference (MRPD) = -34.3 %), with a good agreement (intraclass correlations coefficient (ICC) = 0.77). The global TB in the whole cohort was overestimated by 28.3 %, but patient-level TB was on average (median) underestimated by 1.1 cc (MRPD = -18.4 %) with high variability with a median absolute relative percentage difference (MARPD) = 68.6 %) and poor agreement (intraclass correlation coefficient (ICC) = 0.28).

Conclusions: PARS shows potential for treatment decision support with moderate accuracy in lesion detection and lesion volume estimation, but demonstrates significant variability in TB estimation, highlighting the need for further model refinements before clinical adoption.

Keywords: Autosegmentation; Deep-learning; Metastatic melanoma.

Fig. 1

Examples of the differences between expert and PARS segmentations. A: Substantial underestimation of TB by PARS; B: Good match between experts and PARS; C: Substantial number of false positives and overestimation of TB.

Fig. 2

Precision and recall of PARS model. A: Precision and recall of the PARS model plotted against the PARS probability threshold for different tumor sites. B: Precision–recall curves for each site, illustrating the relationship between precision and recall across all possible probability thresholds.

Fig. 3

Comparison of PARS and expert assessments for three metrics: (A, B, C) the number of lesions, (D, E, F) the lesion size (in cc), and (G, H, I) the total tumor burden (in cc). The top row of boxplots (A, D, G) shows the absolute difference (PARS – Experts), while the middle row (B, E, H) shows the relative percentage error on a symlog scale. In each boxplot, the central line denotes the median, box edges indicate the interquartile range, and whiskers extend to 1.5 × IQR. Outliers are shown as open circles. The bottom row (C, F, I) displays scatter plots of PARS versus Experts, with the diagonal line indicating equality.