ESTRO 2020 Abstract book

S889 ESTRO 2020

We retrospectively analyzed 77 HNSCC patients with pre- treatment CT scans downloaded from The Cancer Imaging Archive (TCIA). Primary tumor was manually delineated on all evaluable artifact-free slices by a single physician and the delineation was independently confirmed by a radiologist. Using the CERR radiomics toolbox, 104 two- dimensional radiomic features were extracted from segmented regions of interest. After stability test, 67 radiomic features were selected for further analysis. Potential subgroups of HNSCC were identified using unsupervised consensus clustering. Immune microenvironmental variables were assessed to investigate biological differences between the resulting radiomic subgroups. A predictive model of CD8+ T-cell fraction was built using random forest regression on radiomic features in a 10-fold cross validation manner. Radiomic clustering results were validated utilizing an external HNSCC cohort (n=83) treated at our institution. Results The tumor subsite consisted of oral cavity (n=38), larynx (n=28), and oropharynx (n=11) with 13 HPV positive tumors. Consensus clustering resulted in two distinct subgroups with 39 and 38 samples in clusters 1 and 2, respectively (Figure 1). A significant difference in tumor subsite between the two radiomic clusters was found with p=0.0096. Cluster 1 was significantly enriched for oropharyngeal tumors (n=10) along with 14 laryngeal and 15 oral cavity tumors. By contrast, cluster 2 was enriched with oral cavity primaries (n=23) along with 14 laryngeal tumors and 1 oropharyngeal tumor. There was a significant difference in HPV status between the two radiomic clusters with p=0.0127. A random forest regression model of CD8+ T-cell fraction was built using 67 radiomic features. A 10-fold cross validation strategy resulted in significant prediction power with R 2 =0.30 (p<0.0001) (Figure 2). A significant association between HPV-positivity and increased CD8+ T-cell fraction was found with p=0.0061. Consensus clustering was performed on the same 67 radiomic features for the validation cohort with 31 oropharyngeal tumors with 27 HPV positivity, 1 laryngeal, and 51 oral cavity tumors, resulting in two separable radiomic clusters with 41 and 42 samples, respectively. A significant difference in tumor subsite between the two radiomic clusters was found with p=1.3×10 -7 . HPV status is not routinely assessed for oral cavity tumors in our institution due to its rare positivity and was not available in the validation cohort. HPV status was associated with two radiomic clusters if all oral cavity cancers were assumed to be HPV-negative (p=4.0x10 -7 ). Conclusion Radiomic features extracted from CT scans were correlated with the tumor immune microenvironment and HPV status in HNSCC, suggesting that radiomics has significant potential for imaging biomarker development.

PO-1551 Deep CNN on PET/CT images for NSCLC automated tumor detection and outcome prediction M. Ibrahim 1 , J. Castelli 2 , C. Cheze Le Rest 1 , O. Acosta 2 , D. Visvikis 1 , R. De Crevoisier 2 , M. Hatt 1 1 University of Western Brittany, LaTIM- INSERM- UMR 1101, Brest, France ; 2 University of Rennes, LtSI-INSERM- UMR 1099, Rennes, France Purpose or Objective Our goal was to externally validate a previously validated workflow based on deep convolutional neural networks (CNN) for non-small cell lung cancer (NSCLC) patients outcome after chemoradiotherapy fully automated prediction from positron emission tomography / computed tomography (PET/CT) images, i.e. without the need for tumor segmentation, as is usually required in radiomics. Material and Methods Two cohorts from two institutions (Poitiers, n=110 and Rennes, n=37) of stage II-III NSCLC patients were used for training/validation and external testing respectively. The workflow consists of 2 deep CNNs. The first is a 3D CNN aiming at primary tumor detection, which was trained on 2 publicly available datasets of CT scans (1397 and 888 patients) with or without lung nodules/tumors (not necessarily NSCLC). This trained tumor detector was then fine-tuned (i.e., transfer learning) then evaluated in the training cohort using the low dose CT from the PET/CT images as input. The goal is then to feed only the part of the image containing the primary tumor to a second CNN, trained specifically for outcome prediction (classifying