ESTRO 38 Abstract book

S1038 ESTRO 38

Material and Methods Twenty patients with stage II, III or IV head and neck squamous cell carcinoma were included in the PREDICT study. Two patients had HPV positive tumors. Treatment consisted of radiotherapy with or without concurrent chemotherapy. All patients underwent MRI prior to and during week 2, 3, 4 and 5 of the radiotherapy treatment. Imaging was obtained with the patient positioned in the radiotherapy mask. Tumor delineation was performed on T2 weighted images for the baseline MRI and each subsequent MRI. Volume changes were determined using these delineations. For the ADC changes the tumor was again delineated on the baseline b=800 s/mm 2 images using a semi-automatic method. This delineation was copied to the corresponding ADC map and ADC maps of subsequent weeks. Median ADC values for each available delineation were extracted. A follow up of at least 3 months was available for all patients. Results During (chemo)radiotherapy tumors generally reduce in size with each passing week. On average the tumors were only 50% of their original size at the end of the third week of treatment. However due to the treatment effects, tumors are increasingly harder to differentiate from nonmalignant tissues in the treatment area. At the end of the fifth week only 20% of the original tumor volume was visible on T2. In one patient the tumor visibly increased in size from the third week onward. This patient had a local recurrence within 3 months after treatment. In total four patients had a recurrence. ADC generally increases during therapy. The fractional change in median ADC between the baseline and third week of treatment (fADC 3 ) was on average 1.25 and the most discriminating between patients with recurrence and with response. A cutoff of 32% increase in ADC at week 3 resulted in a sensitivity of 100% and specificity of 81%. Conclusion During (chemo)radiotherapy, T2 images can be used to measure tumor volume. Generally tumors decrease in size during treatment. A large increase at week 3 however, might predict a recurrence of tumor. Figure 1. Fraction tumor change volume of patients (black), average of all patients with recurrence (red) and average of patients without recurrence (green).

Purpose or Objective Radiomics is a promising tool for identification of new prognostic biomarkers. Radiomic models are often based on single-institution data. However, multi-centric data that are highly heterogeneous due to different scanning protocols reflect better the clinical reality. Robustness studies are crucial to find features independent from e.g. scanner settings. We studied if a CT radiomics overall survival (OS) model trained on multi-centric data with prior robust feature selection can achieve a similar performance as a model on standardized data. Material and Methods Pre-treatment CT data from 121 stage IIIA/N2 NSCLC patients from a prospective Swiss multi-centric randomized trial (SAKK 16/00, neoadj. chemo- or radiochemotherapy prior to surgery) were used to calculate 1404 radiomic features on the primary tumor. Two OS radiomic models were trained on (1) a patient sub- cohort characterized by standardized imaging protocol (native CT, standard kernel, n = 84) and on (2) the entire heterogeneous patient cohort but with pre-selection of robust radiomic features. Robust features were extracted from four distinct robustness studies (contrast, convolution kernel, motion, delineation). Stability measure was the intra-class correlation coefficient (> 0.9 considered stable). Principal component (PC) analysis was performed for feature selection and PCs describing in total 95% of data variance were selected. Features were selected separately for the entire and standardized dataset. The feature with highest correlation to the PCs served as a surrogate for the multivariate Cox model. Finally, backward selection was performed. Model performance was quantified using Concordance Index (CI). 10-fold cross-validation and bootstrap with resampling were used both to verify and compare model performances. Results Robustness studies revealed 113 stable features (n shape = 8, n intensity = 0, n texture = 7, n wavelet = 98). The convolution kernel was the largest influence on the robustness of the radiomic features. The final OS model on the entire non- standardized dataset consisted of four and the model on standardized data of six features (all identified as unstable). The model on standardized imaging data showed significant better prognostic performance compared to the model with robust feature pre-selection based on the entire heterogeneous imaging data (CI = 0.64 and 0.61, p < 0.05, resp.). Conclusion For our prognostic NSCLC radiomic models, image protocol standardization appears superior to using larger but heterogeneous imaging data combined with robust feature selection. EP-1911 Treatment response on MR during radiotherapy in patients with head and neck squamous cell carcinoma. B. Peltenburg 1 , M. Philippens 1 , R. De Bree 2 , C. Terhaard 1 1 UMC Utrecht, Radiotherapy, Utrecht, The Netherlands ; 2 UMC Utrecht, Head and Neck Surgical Oncology, Utrecht, The Netherlands Purpose or Objective In head and neck radiotherapy, early recognition of patients with poor response to treatment is important and might allow for treatment modification. Conventionally, tumor volume changes are used to assess treatment response. Recently, apparent diffusion coefficient (ADC) determined by diffusion weighted magnetic resonance imaging (DW-MRI) has been introduced as a prognostic factor in patients with head and neck squamous cell carcinoma. Aim: To follow treatment response on DWI and T2 weighted images of head and neck tumors.

Figure 2. T2 weighted MRI images of the patient with a local recurrence within 3 months after treatment. A) Pretreament MRI with the tumor in the hypopharynx (white arrow) . B) Week 2 C) Week 3, D) Week 4, E) Week 5 of radiotherapy. F) MRI of local recurrence 3 months after radiotherapy.