ESTRO 2025 - Abstract Book

S964

Clinical – Head & neck

ESTRO 2025

less than 20% fell within this area. A total of 39 radiomic features were extracted from the reirradiated GTV (rGTV) using LIFEx software. Feature selection was performed using Spearman’s correlation and the Wilcoxon test, resulting in three feature combinations for analysis. Following Principal Component Analysis (PCA), the two main components of each combination defined a 2D plane, enabling separation of cases based on point distributions. Classification algorithms (Linear and Quadratic Discriminant Analysis) were then applied to separate “in-field” and “outside” recurrences. Classification probabilities were calculated with a decision threshold specific to each feature combination, and scores were averaged, weighted by the quality of representation of each point within the three 2D planes. The decision threshold was further weighted by the sum of variances explained by the PCA axes. Performance metrics, including AUC, Brier’s score, and Balanced Accuracy (BA), evaluated the model’s efficacy. Results: The 44 patients were split into a training cohort of 23 patients (9 "in-field," 14 "outside") and a testing cohort of 21 patients (11 "in-field," 10 "outside"). The three models (Figure 1) obtained were: combination A, where LDA achieved a BA of 87% in training and 76% in validation with a decision threshold of 0.5; combination B, with a BA of 78% in training and 71% in validation using the same threshold; and combination C, which matched combination B’s BA but applied a Youden’s index threshold of 0.71. These results yielded a weighted average accuracy (Figure 2) of 87% (19/23) in training and 90% (19/21) in validation.

Conclusion: This study identified a radiomic signature from pre-reRT PET scans by calculating a weighted average of three models, enabling accurate assessment of second in-field recurrence risk. If validated in larger cohorts, this signature could enhance treatment planning prior to reRT for recurrent HNC.

Keywords: radiomics, ensemble learning, treatment adaptation