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

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Artificial intelligence (AI) applications in radiation oncology are expanding, yet tumor burden estimation using AI remains under-explored.
• Deep learning models offer potential for improving diagnostic accuracy in cancer imaging.

Purpose of the Study:

• To evaluate the performance of a deep learning model, PET-Assisted Reporting System (PARS), for lesion detection, segmentation, and tumor burden estimation.
• To assess the utility of [18F]-FDG-PET/CT-based AI in patients with metastatic melanoma undergoing immunotherapy.

Main Methods:

• Retrospective analysis of 165 stage IV melanoma patients who underwent [18F]-FDG-PET/CT before immunotherapy.
• Comparison of gross tumor volumes segmented by PARS with manual delineations by radiation oncologists.
• Assessment of lesion detection metrics (precision, recall), individual lesion volume agreement, and overall tumor burden estimation accuracy.

Main Results:

• PARS achieved a recall of 68.9% but a modest precision of 46.8%, with performance varying by lesion location (e.g., lung vs. bone).
• PARS tended to underestimate individual lesion volumes (median relative difference -34.3%) with good agreement (ICC=0.77).
• Overall tumor burden was overestimated globally (28.3%), but patient-level estimation showed underestimation (median -18.4%) with high variability and poor agreement (ICC=0.28).

Conclusions:

• The PARS AI model demonstrates moderate accuracy for lesion detection and volume estimation, suggesting potential for treatment decision support.
• Significant variability in tumor burden estimation necessitates further model development for reliable clinical adoption in metastatic melanoma management.