ESTRO 2024 - Abstract Book

S1210

Clinical - Head & neck

ESTRO 2024

593

Digital Poster

Using [18F]-FDG PET Radiomics to Predict Survival After Reirradiation in Head and Neck Cancer

Arnaud BEDDOK 1,2 , Kira Grogg 2 , Laura Rozenblum 3,2 , Christophe Nioche 1 , Fanny Orhlac 1 , Valentin Calugaru 4 , Gilles Crehange 4 , Helen Shih 5 , Thibault Marin 2 , Georges El Fakhri 2 , Irène Buvat 1 1 Institut Curie, Inserm LITO U1288, Orsay, France. 2 Massachusetts General Hospital, Radiology, Boston, USA. 3 Pitié Salpêtrière, Nuclear Medicine, Paris, France. 4 Institut Curie, Radiation Oncology, Paris, France. 5 Massachusetts General Hospital, Radiation Oncology, Boston, USA

Purpose/Objective:

Curative reirradiation (reRT) is an emerging strategy for treating local recurrence in head and neck cancer (HNC). However, nearly half of the treated patients die within two years post-reRT 1,2 . This study aims to assess the prognostic value of [18F]-FDG PET (PET)-derived radiomic features in predicting two-year overall survival (OS) post reRT.

Material/Methods:

A collaborative study was initiated, comprising 67 patients treated for recurrent HNC, 31 from Institut Curie, Paris, France and 36 at Massachusetts General Hospital, Boston, USA. All patients underwent PET imaging before reRT. The reirradiated gross tumor volume (rGTV) was segmented on the PET images using a SUV > 3. Thirty-nine radiomic features were extracted using the LIFEx software after performing 2x2x2mm spatial resampling and setting a fixed bin size to 0.157 SUV units. Feature selection was guided by Student’s t -tests and correlogram analysis to discriminate between patients alive two years post-reRT versus those who were not. Principal Component Analysis (PCA) was then used on these selected features to build a predictive model for OS. The model’s robustness was evaluated through leave-one-out cross-validation and permutation tests, using the balanced accuracy (BA) figure of merit. Survival curves were generated using Kaplan-Meier (KM) methods.

Results:

Of the 67 initial patients (median age 58, 41 males, 26 females), 12 were excluded from radiomic analysis due to a follow-up time less than 2 years. The primary re-irradiated sites were nasopharynx (26.9%) and cervical nodes (23.9%), with squamous cell carcinoma accounting for 71.6% of cases. Intensity-Modulated Radiation Therapy (IMRT) was the most used reRT technique, at 49.4%, and the median radiotherapy dose was 66 Gy ( Table 1 ). One, two, and four-year OS rates at 73.5% [CI95%: 63.5 – 85.2%], 53.3% [CI95%: 41.8 – 67.9%], and 24% [CI95%: 13.6 – 42.2%], respectively. The radiomic analysis identified five textural features that significantly varied between OS ≥ 2 years and OS < 2 years: GLRLM_GLNU, GLRLM_RLNU, NGTDM_Coarseness, NGTDM_Busyness, and GLSZM_GLNU. Three of these features — GLRLM_GLNU, GLSZM_GLNU, and NGTDM_Busyness — were not correlated one with the other (Spearman coefficient < 0.70). PCA-based classification using the value of these three features identified two OS groups in a 2D space (p < 0.02) (Figure 1 ). Based on the first principal component only, patients with an OS greater