A Multichannel Markov Random Field Approach for Automated Segmentation of Breast Cancer Tumor in DCE-MRI Data Using Kinetic Observation Model Ahmed B. Ashraf, Sara Gavenonis, Dania Daye, Carolyn Mies, Michael Feldman, Mark Rosen, and Despina Kontos The University of Pennsylvania, Philadelphia, PA 19104, USA Ahmed.Ashraf@uphs.upenn.edu Abstract. We present a multichannel extension of Markov random fields (MRFs) for incorporating multiple feature streams in the MRF model. We prove that for making inference queries, any multichannel MRF can be reduced to a single channel MRF provided features in dif- ferent channels are conditionally independent given the hidden variable. Using this result we incorporate kinetic feature maps derived from breast DCE MRI into the observation model of MRF for tumor segmentation. Our algorithm achieves an ROC AUC of 0.97 for tumor segmentation. We present a comparison against the commonly used approach of fuzzy C-means (FCM) and the more recent method of running FCM on en- hancement variance features (FCM-VES). These previous methods give a lower AUC of 0.86 and 0.60 respectively, indicating the superiority of our algorithm. Finally, we investigate the effect of superior segmentation on predicting breast cancer recurrence using kinetic DCE MRI features from the segmented tumor regions. A linear prediction model shows sig- nificant prediction improvement when segmenting the tumor using the proposed method, yielding a correlation coefficient r =0.78 (p< 0.05) to validated cancer recurrence probabilities, compared to 0.63 and 0.45 when using FCM and FCM-VES respectively. Keywords: Breast DCE MRI, breast tumor segmentation, tumor char- acterization, breast cancer recurrence prediction. 1 Introduction Crucial to the performance of a feature extraction and image classification sys- tem is the availability of a reliable segmentation approach for the object of interest (e.g. tumor). In most medical imaging scenarios the automation of this step is particularly important because of the large amount of images to be ana- lyzed. This makes the manual segmentation approach tedious and prohibitively expensive. As a result, a lot of research has been done in the medical imaging community for improving the quality of automated segmentation. Specifically, G. Fichtinger, A. Martel, and T. Peters (Eds.): MICCAI 2011, Part III, LNCS 6893, pp. 546–553, 2011. c  Springer-Verlag Berlin Heidelberg 2011