Abstract No. 241 Fuzzy C-Means Segmentation of Hyperintense Metastatic Lesions in Hepatocyte-Avid Contrast T1-weighted DCE-MRI Studies

At 90%, the liver is the main site of secondary cancer deposition in metastatic uveal melanoma (MUM) and is a leading cause of mortality. Dynamic contrast-enhanced (DCE) MRI studies using hepatocyte-avid contrast are routinely done as the contrast agent is absorbed and retained by hepatocytes but not by the in situ metastasis. The high degree of spatial resolution of lesion volume, as well as rich temporal information from contrast kinetics in DCE-MRI, has been shown to have important diagnostic value in other oncologic systems, such as breast cancer outcomes. The purpose of this study is to evaluate the performance of automatic fuzzy C-means segmentation of tumor lesions on DCE-MRI studies of liver metastatic ocular melanoma cases and explore the associations between computer-extracted radiomic features and patient overall survival (OS). Retrospective dataset of DCE-MR studies from 201 patients treated at the University Hospital. Each study is a T1-weighted hepatocyte-avid DCE axial sequence taken immediately prior to therapeutic intervention by Interventional Radiology. We have acquired expert radiologist annotation of 642 unique metastatic lesions from these images. Metastatic lesions were automatically segmented using a fuzzy C-means algorithm with bounding-box initialization. Following segmentation, radiomic features such as morphology, texture, and contrast kinetics were extracted from each lesion MRI sequence. Sørensen–Dice coefficient was used to evaluate the overlap between automatic segmentation and expert manual segmentation of metastatic lesions. In a preliminary set of DCE images, fuzzy C-means algorithm shows promising results in the automatic segmentation of hyperintense lesions on DCE-MRI images with a DICE coefficient of 0.70 (n=3). In the coming months, we will be evaluating the segmentation performance of this algorithm on the full annotated set of lesions as well as the generalizability of the method to hypointense metastases. Fuzzy C-means algorithm shows potential as an automatic segmentation method for metastatic lesions on hepatocyte-avid DCE-MR images. There is great utility in a tool that can automatically segment and extract radiomic features from metastatic lesions; allowing us to better study these patients in the hope of ultimately improving long-term outcomes.