Abstract ED01-02: Imaging ‘omics in cancer detection

Abstract

Abstract “Radiomics” is the process of extracting quantitative descriptive features from biomedical images in high throughput. These image features are thus mineable data that can be used to parse against medical outcomes or gene expression data. A long term goal is to develop models using these data that will enable improved detection, diagnosis and therapy prediction. Two projects in lung cancer are currently underway. The first focuses on non-small cell lung cancer (NSCLC) imaged with standard-of-care PET/CT and CT images with 3–5mm resolution. For CT scans, the Radiomic dataset includes 219 three-dimensional and 102 two-dimensional features from each CT images. Data have been generated for over 1,000 lung cancer patients who have undergone surgical biopsy. Of these, 276 samples have been analyzed for gene expression using Affymetrix platforms and analyses are underway to determine relationships between image features, clinical outcomes and gene expression. In developing this technology, we have improved methods for automated delineation of lung lesions. This uses a multi-seed point approach to initiate region growing algorithm to generate an ensemble segmentation from multiple regions. This has been evaluated on a set of lung tumor datasets from the Moffitt Cancer Center thoracic oncology database. The similarity index (SI) was calculated as the ratio of the union over the intersection across multiple segmentations from the same lesion. In this metric, a SI of 100% reports identical results from every initial seed. For our data set, the average SI was > 93% in 121/129 patients with 20 different start seed points for each case. This data set has been uploaded to CABIG using the NBIA platform and is available as a “segmentation challenge”. In our own data set, if PET scans are available, we are using centroids of the maximal SUV to automatically detect seed points. Another important step in the qualification process is to assess the quality of the data through co-variance, test-retest reproducibility and biological ranges of these features. The full feature sets have been analyzed for covariance, and a number of redundant sets of features have been identified, which allowed a reduction in the dimensionality of feature space. Test-retest analyses have baseline and follow-up thoracic CT scans (slice thickness 1.25mm) that were obtained within 15 minutes on 32 patients with NSCLC. All patients had primary pulmonary tumor of 1 cm or larger. Following segmentation, 190 3D quantitative features were extracted and the reproducibility of these tumor features in the two scans performed on each patient was assessed by calculating the concordance correlation coefficient (CCC). In general, the intra-patient reproducibility of all features was high, 102/190 showed excellent reproducibility (CCC>0.75). Notably, the inter-patient biological ranges for individual features were highly variable. Additionally, a co-variance matrix of features identified several redundancies in the feature set which could be combined into a single variable. Combining inter-scan variance, biological range and co-variance, we have reduced the total number of features from 190 to a set of ∼20 that may be the most informative. Comparisons of quantitative imaging feature data with patient outcome have begun with an approach to classify Adenocarcinoma, Squamous-cell Carcinoma and Bronchioalveolar Carcinoma (BAC). Classifiers including decision trees and support vector machines are used along with feature selection techniques (Wrappers and Relief-F) to build effective models for tumor classification. Results show that over the large feature space for 3D features it is possible to recognize tumor classes with over 76% accuracy, showing new features may be of help. The results indicate that image features on their own can be used to identify tumor classes with relatively high accuracy. The second project has just been initiated to use the Radiomics approach to extract features from screening CTs. This is using the ACRIN-sponsored National Lung Screening Trial, NLST, database of ca. 50,000 individuals who were screened with CT or planar X-ray. The results of this study are now being published and have shown a significant survival benefit in the CT-screened population. A concern with this approach is the large number of false positives as well as nagging false negatives, i.e. patients who presented with late-stage cancer following a “negative” exam. The Radiomics approach will test the hypothesis that quantitative texture and other features will add to predictive power and result in improved sensitivity and specificity. Because the data set is large, significant effort is being expended to automate the 3-D segmentations and identify regions of interest within the lung fields. Citation Information: Cancer Prev Res 2011;4(10 Suppl):ED01-02.