TU‐CD‐BRB‐11: A Spatiotemporal Image Phenotyping Pipeline for Radiogenomic Profiling of Breast Cancer with DCE MRI

Abstract

Purpose:Objective and quantitative radiophenotyping from cancer patients is a core requirement of radiogenomic research. We present our quantitative image phenotyping pipeline which is tuned to extract spatiotemporal tumor traits from DCE breast MRI and initial results of radiogenomic association between spatiotemporal imaging features and RNA expression in breast cancerMethods:DCE MRI with 5 enhancement phases of 61 breast cancer patients were analyzed. A unified energy functional was defined for a joint shape and temporal motion estimation which incorporates intensity variation within a region along with kinetic signature. A level set evolution was carried out under the guidance of the unified energy functional to generate tumor probability maps, which is then thresholded to produce tumor masks. Among them, the observer selected the largest one. A total of 47 features were extracted from the segmented tumor comprising geometric, statistical, and spatiotemporal features. The whole procedure was carried out independently by two observers. Features from 19 cases were used as a training set to discover a prognostic imaging biomarker indicative of poor metastasis‐free survival, and the remaining 39 cases were used for validation.In addition, association between the predictive imaging biomarker and genomic profiles were assessed using RNA expression data of these matched patients obtained with next‐generation RNA sequencing and RT‐PCR.Results:All imaging features from two independent analysis were highly consistent (r > 0.78, p<0.001). Among them, a spatiotemporal feature (enhancing rim fraction score, ERF) was shown to be strongly predictive of early metastasis (P =.009; hazard ratio,16.3; 95% confidence interval: 1.99). Radiogenomic analysis revealed five named lncRNAs significantly associated with high ERF score.Conclusion:A quantitative image phenotyping pipeline was developed which provides objective spatiotemporal features and could be successfully utilized in radiogenomic profiling of breast cancer.