Tumor probability mapping of fluorescence intensity during image-guided surgery of head and neck cancer

Abstract Achieving negative surgical margins remains a critical determinant of local recurrence and survival in head and neck cancer (HNC) surgery. Current intraoperative margin assessment techniques, including frozen section analysis, suffer from sampling errors and procedural delays. Tumor-targeted fluorescence imaging offers real-time tumor visualization but lacks standardized quantitative approaches for clinical decision-making. We developed a Tumor Probability Mapping (TPM) framework using panitumumab-IRDye800 fluorescence imaging in 16 HNC patients. Ex vivo specimens and gross tissue sections were imaged using near-infrared fluorescence systems. A total of 5,442 regions of interest (ROIs) were manually distributed across fluorescence images of gross specimen sections validated by histopathology. Signal-to-background ratios (SBR) were calculated and used to train the following predictive models: generalized linear model fit standard logistic regression (MATLAB, glmfit), standard logistic regression (R, LOG), mixed-effects logistic regression (GLMER), and Bayesian mixed-effects regression (BRMS). Model performance was evaluated using receiver operating characteristic and area under the curve (ROC-AUC) analysis, sensitivity, specificity, along with beta-calibration and model fit. All models demonstrated excellent (> 90%) discriminative ability between tumor and normal tissue. The glmfit model, selected for clinical implementation, achieved 95.8% accuracy, 90.8% sensitivity, 98.8% specificity, and an AUC of 0.989 on test data. The final TPM algorithm provides real-time probability assessment of tumor presence based on fluorescence intensity quantified by histopathology validated historical data. TPM represents a significant advancement in fluorescence-guided surgery by converting qualitative fluorescence signals into quantitative probability assessments validated against histopathology. This approach provides surgeons with standardized, real-time tumor probability information that extends beyond qualitative assessments and/or binary threshold determinations, potentially improving surgical outcomes by enhancing margin assessment and reducing local recurrence rates.