Meeting Banner
Abstract #3363

A 3D Shape and Textural Classification Tool for Identifying Malignant Breast Cancer

Rebecca E. Thornhill1, Greg O. Cron2, Kevin Ibach1, Shilpa Lad1, Mark E. Schweitzer1, Jean Seely1

1Medical Imaging, The Ottawa Hospital, Ottawa, Ontario, Canada; 2Medical Imaging, Ottawa Hospital Research Institute, Ottawa, Ontario, Canada

While breast MRI has shown tremendous promise for characterizing breast cancers, its specificity has been limited by reliance on tumor shape. Many tumors will exhibit discrete areas of high perfusion or vascular leakiness. These hot spots could yield important information that would be obscured by reporting the average tumor Ktrans. In this study, we have identified a potential recipe for predicting malignant breast cancer comprising of shape and textural features. While textural features appear to provide good specificity and modest sensitivity, the converse was true for shape-based models. With further optimization, this approach may improve accuracy compared to conventional MRI.

Abstract PDF Presentation Video

Keywords

accuracy achieve aggressive aided although among appear appeared array assessed automatic basic benign bergen biometrics bliss branch breast cancer cavities class classification coefficient coefficients coil combined combining comp comprising computed computer contrast contributing conventionally converse corrected criteria curves datasets dedicated defined descriptors designating diagnostic discrete discriminative downstream dynamic eccentricity emphasis enhanced entire entropy evidence exhibit feature features fisher gadolinium generated good gradient gray highest identified identify identifying improved included injected interventions length lesion likely logistic malignancy malignant manually many maps margins mark matrix median model models modest morbidity morphological obscured occurrence ongoing outcomes paucity performance pixel potential predicting preliminary process prognostic prophylactic proportional proven radiation receive recent regression reporting resolution retrospectively revealed risk roundness runs saturation sensitivity serious shape short signify slice slices software specificity spots studied studies sufficient suggests summary table terms textural texture tool tools topology trans true tumor tumors undergoing uniformity variance versa version vice whole yang yield