ESTRO 2020 Abstract book

S891 ESTRO 2020

Conclusion Radiomic features from CE-CT could help in the selection of patients for a laryngeal preservation strategy. Statistical harmonization based on ComBat seems to improve the predictive value of radiomic features extracted in such a highly heterogeneous multicentric setting. These findings now require evaluation in an external cohort PO-1554 Radiogenomics: A ‘Virtual Biopsy’ in Non- small Cell Lung Cancer? H. SAXBY 1 , H. Wang 2 , V. Ezhil 1 , P. Evans 2 , M. Halling- Brown 3 , I. Phillips 4 , V. Prakash 5 , A. Nisbet 6 1 Royal Surrey County Hospital, Oncology, Guildford, United Kingdom ; 2 University of Surrey, Centre for Vision- Speech and Signalling, Guildford, United Kingdom ; 3 Royal Surrey County Hospital, Scientific Computing, Guildford, United Kingdom ; 4 Great Western Hospital, Oncology, Edinburgh, United Kingdom ; 5 Royal Surrey County Hospital, Nuclear Medicine, Guildford, United Kingdom ; 6 University College London, Physics, London, United Kingdom Purpose or Objective Lung cancer is the second most common malignancy in the UK. Its treatment depends on multiple factors including the histology and genes expressed by the tumour. CT scans play a very important role in the diagnosis and management of patients with lung cancer. The emerging field of radiogenomics aims to extract further quantitative information from imaging with regards to the tumour genotype. This pilot study aims to explore the relationship of CT texture features between the two main histological subtypes of non-small cell lung cancer (NSCLC); adenocarcinoma and squamous cell carcinoma. Material and Methods The lung tumour was segmented on the diagnostic CT scans of 10 patients with biopsy proven stage I NSCLC adenocarcinoma and 10 patients with biopsy proven stage I NSCLC squamous cell carcinoma. The scans were anonymised and exported to a programme developed in house [Wang et al] which was then used to extract 43 common texture features from the CT defined tumours. 3 features were first order and 40 were higher order. A p- value was calculated using a two sample t-test for each feature. Results 9 of the 43 texture features showed a statistically significant difference between the adenocarcinoma and squamous cell carcinoma lung tumours, demonstrating the potential of CT texture to be employed to differentiate between histological subtypes Figure 1: The texture features that demonstrated a statistically significant difference

Purpose or Objective To develop a radiomic model predicting non-response to induction chemotherapy in laryngeal cancers from pre- therapeutic contrast-enhanced computed tomography (CE-CT). Material and Methods One hundred and five patients eligible for laryngeal preservation chemotherapy from June 2008 to January 2018 were considered in 5 centres. Ninety-eight patients were analyzable for the endpoint defined as non-response to therapy according to clinical exam. Primary tumor was manually delineated on the CE-CT images. IBSI-compliant radiomic features were extracted. Features harmonization was performed using ComBat after unsupervised hierarchical clustering to identify two groups of patients with similar features distributions, due to the very high heterogeneity of CE-CT acquisitions settings and reconstructions parameters. The cohort was split into training/validation (n=66) and testing (n=32) sets. Receiver operating characteristics curves were used to evaluate the predictive value of radiomic features and clinical variables in the training/validation set. Features were selected based on an area under the curve (AUC) ≥ 0.650, lack of redundancy and a clinically-relevant specificity (sp) above 60% after optimal threshold identification with the Youden index.

Results None of the patient- and/or tumor-related variables were significantly correlated with non-response. Without harmonization, none of the CE-CT radiomic features identified in the training/validation set had predictive power in the testing set. After ComBat harmonization, Zone Size Percentage GLZSM was significantly correlated with non-response to chemotherapy in the training set (AUC = 0.67, Se = 70%, Sp = 64%, p=0.04) and obtained a satisfactory performance in the validation set (Se = 80%, Sp = 67%, p=0.03).

Grey Level Co- occurren ce Matrix

p values 0.018 0.0180.0 23 0.009 0.009 0.031 0.007 0.0360.0 36

Energy Entropy Homogeneity

Short

run

emphasis

Grey Level Run

Run non-uniformity High grey level run emphasis Short run high grey level emphasis Long run low grey level emphasis Run level variance length

Length Matrix

Conclusion The purpose of this work is to test the hypothesis that patients could avoid tissue biopsies if we can determine the histology and genes expressed by the tumour from imaging alone. These biopsies are not always technically possible and can cause significant morbidity in this often