A Lesion-adaptive Segmentation Approach for Tumor Delineation on FDG PET/CT in Diffuse Large B-cell Lymphoma Patients

Area of Science:

• Nuclear Medicine
• Oncology
• Radiomics

Background:

• Standard SUV4.0 thresholding for metabolic tumor volume (MTV) in diffuse large B-cell lymphoma (DLBCL) may be unreliable for interim and end-of-treatment (EoT) [¹⁸F]FDG PET scans due to heterogeneous residual uptake and lower contrast.
• Accurate MTV assessment is crucial for treatment response evaluation in DLBCL.

Purpose of the Study:

• To develop and validate a lesion-adaptive decision rule for selecting the optimal [¹⁸F]FDG PET segmentation method based on lesion characteristics and treatment phase in DLBCL.
• To compare the performance of this decision rule with machine learning (ML) based selection models.

Main Methods:

• 598 lesions from 33 DLBCL patients were segmented using six semi-automated methods (SUV2.5, SUV4.0, 41%max, A50peak, MV2, MV3) at baseline, interim, and EoT PET/CT.
• Segmentation quality was rated by expert observers, and the influence of lesion features (SUVpeak, TBRpeak, SUVbg) and treatment phase was assessed.
• A lesion-adaptive decision rule was derived by evaluating over six million rule-based combinations of key features.

Main Results:

• A simple decision rule (SUV4.0 if SUVpeak > 8, MV3 if SUVbg > 0.8, else MV2) achieved 0.82 lesion-wise accuracy for preferred method selection, matching best ML models.
• The decision rule improved lesion-level MTV agreement (ρ=0.85) and reduced MTV deviation (>10%) compared to SUV4.0 alone (ρ=0.82, 63.5% deviation).
• Total-MTV agreement was high across all methods, with modest gains for the decision rule at interim and EoT PET.

Conclusions:

• A straightforward decision-rule approach using SUVpeak and SUVbg effectively selects the optimal segmentation method for individual DLBCL lesions across treatment phases.
• This method offers greater simplicity and clinical applicability than ML models, addressing current segmentation gaps for interim and EoT PET where SUV4.0 may be suboptimal.
• The decision rule enhances the reliability of MTV assessment in DLBCL, particularly when standard thresholding methods are less appropriate.