CT Texture Analysis (CTTA): Developing a Diagnostic Imaging Biomarker for KRAS Mutation in Metastatic Colon Cancer
AffiliationThe University of Arizona College of Medicine - Phoenix
MetadataShow full item record
PublisherThe University of Arizona.
DescriptionA Thesis submitted to The University of Arizona College of Medicine - Phoenix in partial fulfillment of the requirements for the Degree of Doctor of Medicine.
AbstractPURPOSE To evaluate multi-parametric modeling on imaging textures from contrast-enhanced, multiphasic computed tomography (CT) for identification of Kirsten rat sarcoma (KRAS) gene mutations in metastatic colon cancer to the liver. METHOD AND MATERIALS This retrospective study included 99 patients diagnosed histologically with colon cancer: 51 KRAS wild-type and 48 KRAS gene mutation. Matched-size regions of interest (ROIs) were drawn over viable tumor and unaffected background liver on multiphase CT. Paired ROIs were spatially rescaled, intensity-normalized, and then analyzed using 3 Texture Algorithms: GLCM, LBP, and Gabor. Feature selection method was based on KNN classifier and DEFS (Differential Evolution-based Feature Selection). For each of the 30 independent experiments, patients were randomly allocated into training (n = 79) and testing (n = 20) datasets to develop predictive models for KRAS gene mutation. Classification models were generated based on: 1) All features; and 2) Selected features as per DEFS. RESULTS Predictive models utilizing all 56 features (13 GLCM, 26 LBP, and 14 Gabor) resulted in an average accuracy/sensitivity/specificity of 61/54/62%; ranging from a single best model (80/80/90%) to a single worst model (35/20/20%). Predictive models utilizing a DEFS optimized 3-feature subset resulted in average accuracy/sensitivity/specificity of 89/80/84%; ranging from a single best model (95/92/96%) to a single worst model (80/68/68%). Among the three texture algorithms, LBP provided better discriminatory power compared to GLCM and Gabor. CONCLUSION Utilizing advanced analytics with machine learning techniques (CTTA and DEFS selection analysis), multi-textural data obtained from conventional, multiphase CT images has the capability to detect a therapeutically relevant genetic aberration (KRAS mutation) in metastatic colon cancer with high accuracy, sensitivity and specificity.