ESTRO 38 Abstract book

S1111 ESTRO 38

Purpose or Objective There are multiple ways to improve radiation therapy for head and neck cancer. One is by dose painting, using the baseline FDG PET. Another is adaptive response assessment, using a midtreatment FDG PET scan. Good predictability of the midtreatment PET is a reason to pursue baseline dose painting. Contrary, if the midtreatment PET does add substantial new information, adaptive response assessment would need further investigation. Therefore, we evaluated the predictability of the midtreatment PET scan, using pretreatment information. Material and Methods We analyzed 39 patients with squamous cell cancer of the oropharynx, oral cavity and hypopharynx. Patients were treated with radiotherapy with curative intent, in combination with cisplatin or Cetuximab. Pretreatment imaging consisted of planning CT, MRI and FDG PET/CT. Patients received a second FDG PET and CT during treatment. All scans were deformable registered to the midtreatment CT. A neural network was formed for a voxel based prediction of the midtreatment FDG PET within the gross tumor volume. We used the mean square error as a loss function. The input were the imaging modalities, dose and the marked variables in table 1. A 6- fold cross-validation was used to create the artificial PET scans (example in figure 1). We compared the real and artificial midtreatment SUV max , SUV mean and the area encompassed by 50% of the SUV max (SUV 50% ). The dice coefficient between the 2 SUV 50% segmentations was calculated. Last, we expanded the artificial SUV 50% or lowered the segmentation threshold to encompass the entire real SUV 50% .

n

Age (years)

average

59

range

35-69

T-stage*

2 3 4 0 1 2 3

1

14 24

N-stage*

6 7

24

2

Site

Oropharynx 33 Oral cavity 4 Hypopharynx 2

HPV-status*

Negative Positive Cisplatin

15 24

Concurrent systemic therapy*

33 Cetuximab 6

Days between PET and start RT* average

13

range

7-28

Fractions until midtreatment PET average

8.5

range 7-11 The variables marked * are included in the neural network. Conclusion This pilot study indicates that midtreatment PET scans can be partially predicted. Model improvements through increased sample size and modified loss functions are subject of further studies. EP-2026 Diffusion weighted textural differences between p16 positive and negative oropharyngeal carcinoma S. Deschuymer 1 , V. Vandecaveye 2 , F. De Keyzer 2 , S. Nuyts 1 1 University Hospital Gasthuisberg, Radiotherapy, Leuven, Belgium ; 2 University Hospital Gasthuisberg, Radiology, Leuven, Belgium Purpose or Objective To explore the feasibility and diagnostic performance of radiomic features (RF) based on diffusion weighted (DW) MRI in differentiating among p16 positive and negative oropharyngeal carcinoma (OPC). Material and Methods Sixty-six patients with histologically proven OPC were prospectively analysed. OPC were considered p16 positive if more than 70% diffuse nuclear and cytoplasmic immunohistochemistry staining was present. 1.5 T MRI with echo-planar DW sequences at 6 b-values (0, 50, 100, 500, 750 and 1000 s/mm²) were acquired before treatment. The region of interest (ROI) encompassing the entire primary tumor volume was manually drawn on the apparent diffusion coefficient (ADC) map by an experienced head and neck radiologist. One hundred and five RF (Table 1), including shape, size, first-order-histogram and textural analysis (TA), were extracted from the ROI with Pyradiomics software and were compared between p16 positive and p16 negative OPC with the two-tailed unpaired Student t-test. The significance threshold was set at a p-value of <0.05. In addition, receiver operating characteristic (ROC) curves were generated to determine the discrimination performance and the optimal cut-off value of the RF. Results Fourteen (21%) tumors were p16 positive. A total of 34 RF were significantly different including volumetric parameters with smaller and more sphere shaped tumors

Results Patient and tumor characteristics are depicted in table 1. The artificial PET underestimated the real midtreatment SUV max , the median difference was 3.4 points (interquartile range 4.1 points). The artificial PET underestimated the real SUV max more when the real SUV max increased (Pearson’s r = 0.94). The predicted SUV mean was also lower, with a median difference of 2.2 points (interquartile range 2.7 points). The artificial SUV 50% segmentations were larger, the median volume was 12.3 cc (interquartile range 15.1) vs 4.6 cc (interquartile range 4.7) for the real SUV 50% . The median dice coefficient between the segmentations was 0.53 (interquartile range 0.18). To encompass the real SUV 50% , the SUV threshold of the artificial scans had to be lowered till SUV 38% (interquartile range of 15%) or a median expansion was needed of 3 mm (interquartile range 2 mm).