Identification of Texture Features from FDG PET images that are associated with Survival Outcomes in Non-small Cell Lung Cancer


Session type:


Mubarik Arshad1,Andrew Thornton2,Nicola Rodgers2,Andrew Scarsbrook3,Garry McDermott3,Gary Cook4,Sue Chua5,Richard O’Connor6,Tara Barwick7,Andrea Rockall7,Eric Aboagye7
1Imperial College London and Imperial College Healthcare NHS Trust,2Imperial College London Cancer Imaging Centre and Department of Nuclear Medicine,3Department of Nuclear Medicine, Level 1, Bexley Wing, St James’s University Hospital, Beckett Street, Leeds, LS9 7TF,4St. Thomas’ Hospital Department of Nuclear Medicine, Westminster Bridge Rd, London SE1 7EH.,5The Royal Marsden Hospital Department of Nuclear Medicine, Downs Rd, Sutton SM2 5PT,6Nottingham University Hospital Department of Nuclear Medicine, Queen's Medical Centre, Derby Rd, Nottingham NG7 2UH.,7Imperial College London Cancer Imaging Centre and Department of Nuclear Medicine, Hammersmith Hospital, Du Cane Road, London W12 0HS



Tumour features derived from image texture represent relationships between grey-scale voxel intensity at local or global level, reflecting underlying tumour heterogeneity. The benefits of using imaging biomarkers over biopsy include its non-invasive nature, whole lesion analysis and ease of temporal follow-up (1). The ability to quantify heterogeneity allows user independency and may guide personalised treatment. The purpose of this study was to identify pre-treatment 2-deoxy-2-(18F)-fluoro-D-glucose PET/CT (FDG-PET/CT) texture features in primary non-small cell lung cancer (NSCLC) as predictors of survival.


This retrospective multi-centre study from 6 centres evaluated 340 subjects with NSCLC (Stage I-III) who underwent FDG-PET/CT prior to radical radiotherapy treatment. Basic PET-derived parameters were documented selecting a minimum tumour volume of 5ml. Primary tumour segmentation was performed using a semi-automated adaptive threshold method (2). 660 texture features, including local, regional, global, fractal and wavelet techniques were extracted for 7 different grey levels using TexLAB v2. The dataset was randomly split into three sets - for independent training, validation and test. Lasso regression analysis (3) was used for feature selection, radiomic signature building and overall survival prediction.


Of 340 patients, 261 died within the follow-up period (median 22 (range 0-85) months). The most robust composite radiomics feature-set comprised of 5 texture attributes derived from 3 different wavelet transformations at 32 grey-level, and was significantly associated with survival in the test set (p=0.008; HR 1.4 with stage). This feature-set was independent of established prognostic markers, such as tumour stage (p=0.012, HR 0.644) and was stable across data from different PET/CT scanners within the collaborating centres.


A radiomics feature-set in NSCLC that predicts survival has been defined in a large multi-centre patient cohort. The use of PET scans from different centres demonstrates the robustness of the technique.