Files
Abstract
Background/Objectives: While deep learning-based AI algorithms have been shown to perform well for OPSCC tumor segmentation, the relative value of contrast-enhanced versus T2-weighted sequences for automated segmentation has not been systematically evaluated. In this study, we compared the sequence-specific deep learning performance on contrast-enhanced T1-weighted fat-suppressed and T2-weighted fat-suppressed MRI in HPV-positive OPSCC.
Methods: Pretreatment MRI from 39 patients with paired sequences from a single center were retrospectively analyzed. OPSCC primary tumors were manually segmented using both sequences, which served as the ground truth. Three sequence-specific configurations were evaluated: contrast-enhanced (CE), T2-only, and combined CE + T2. Quantitative evaluation was carried out on aggregated out-of-fold predictions using Dice score (primary), Surface-Dice@2mm (secondary), and other boundary and volumetric metrics, and paired comparisons (combined vs. T2-only; CE-only vs. T2-only) were performed using an exact Wilcoxon signed-rank test. Qualitative evaluation was performed on 4-point ordinal acceptability ratings recorded using a blind reader study, and the ratings were compared using the exact Wilcoxon signed-rank test (pairwise) and dichotomized acceptability using the McNemar test.
Results: Median Dice was comparable across configurations (0.63 for CE + T2, 0.60 for T2-only, and 0.55 for CE-only). Median Surface-Dice@2mm was highest for the combined configuration (0.62), followed by CE-only (0.6) and T2-only (0.57). Median ASSD were 2.71, 2.98, and 2.98 mm, and median HD95 were 11.39, 15.0, and 11.3 mm for combined, CE, and T2, respectively. The median GTV differences (−1.31, −1.29, and −1.49 mL for combined, T2, and CE, respectively) showed a slight bias toward under-segmentation across all configurations. No significant differences in Dice scores were observed for combined vs. T2 (p = 0.11) or contrast-enhanced vs. T2-only (p = 0.98). Similarly, qualitative analysis also showed no evidence of performance difference for ratings and acceptability rates across sequence configurations (paired Wilcoxon, p ≥ 0.35; McNemar, p = 1.00).
Conclusions: In this single-center study, the segmentation performance using non-contrast sequences was comparable to that using both contrast-enhanced and non-contrast sequences. The drop in performance when the contrast-enhanced sequences were excluded from the combination was not significant. These findings justify multi-center validation to support the feasibility of contrast-sparing automated primary OPSCC segmentation when use of contrast agents is contraindicated.