Clinical Evaluation of Deep Learning for Tumor Delineation on ¹⁸F-FDG PET/CT of Head and Neck Cancer

Summary of study. Causes of exclusion were data incompleteness and failure to pass visual validation. No external scans were excluded to these criteria. Model training used 805 patients (835 scans). Each scan represented unique patient in steps 2, 3, and 4.

Study design comparing expert IOV and AI-to-expert variability. Delineations are exemplified on random patient’s axial ¹⁸F-FDG PET/CT intravenous contrast scan slice.

(A) Comparison of 5 implemented methods trained on 196 patient scans based on DSC. Values above boxes are mean followed by 95% CI in parentheses, with P values below. nnU-Net achieved highest DSC and was further analyzed (denoted AI). (B) Paired comparison of AI-to-expert variability and expert IOV on 64 independent internal test scans. (C) Comparison of AI-to-expert variability on 196 internal (same as nnU-Net in A) and 125 external patients. All 3 comparisons used expert-delineated tumor volumes as reference. Values above boxes in B and C are mean difference followed by 95% CI in parentheses, with P values below. Rhombus shape indicates mean value, and central line represents median. Boxes enclose interquartile range. Whiskers extend to most significant measurement no further than 1.5 × interquartile range from hinge. Data beyond whiskers are plotted individually. Notch roughly represents 95% CI around median. (D) DSC, F1 score (F1), and Hausdorff distance (HD) summary statistics in mean ± SD. Hausdorff distance is undefined when expert or AI includes no volume. Hence, numbers marked with *, **, ***, and **** were based on n = 195, 186, 61, and 63, respectively.

Clinical scan delineated by expert (reference) and by AI, along with AI-to-expert agreement (DSC, 0.92). Shown are axial images of 50-y-old man with HNC of rhinopharynx. ¹⁸F-FDG PET/CT with intravenous contrast agent showed greatly increased activity corresponding to large tumor process in right rhinopharynx, crossing midline and growing frontally into cavum nasi on right, intracranially on right, medially in fossa media, and along dura laterally. In addition, multiple lymph nodes in neck had greatly increased activity bilaterally. AI correctly avoided including physiologically active areas such as saliva, metal artifact–induced activity, nose tip, brain, and optic nerve. HU = Hounsfield units.