LIVER METASTASIS EARLY DETECTION USING FMRI BASED STATISTICAL MODEL Moti Freiman 1 , Yifat Edrei 2,3 , Eitan Gross 4 , Leo Joskowicz 1 , Rinat Abramovitch 2,3 1 School of Eng. and Computer Science, The Hebrew Univ. of Jerusalem, Israel. 2 The G. Savad Inst. for Gene Therapy, Hadassah Hebrew University Medical Center, Jerusalem, Israel. 3 MRI/MRS lab HBRC, Hadassah Hebrew University Medical Center, Jerusalem, Israel. 4 Pediatric Surgery, Hadassah Hebrew University Medical Center, Jerusalem, Israel. Email: freiman@cs.huji.ac.il ABSTRACT We present a novel method for computer aided early detec- tion of liver metastases. The method used fMRI-based statis- tical modeling to characterize colorectal hepatic metastases and follow their early hemodynamical changes. Changes in hepatic hemodynamics were evaluated from T 2 *-W fMRI im- ages acquired during the breathing of air, air-CO 2 , and car- bogen. A classification model was built to differentiate be- tween metastatic and healthy liver tissue. The model was constructed from 128 validated fMRI samples of metastatic and healthy mice liver tissue using histogram-based features and SVM classification engine. The model was subsequently tested with a set of 32 early, non-validated fMRI samples. Our model yielded an accuracy of 84.38% with 80% precision. Index Terms— Liver metastasis, fMRI analysis, early de- tection, statistical analysis, computer-aided diagnosis 1. INTRODUCTION The liver is the second most commonly involved organ in metastatic disease, after the lymph nodes. It is the most com- mon site of visceral metastases for colorectal carcinoma pa- tients, and hepatic metastases are a frequent clinical compli- cation. Some focal lesions may be surgically resectable or treated by means of ablation techniques. Since liver func- tion tests in patients with liver metastases tend to be insensi- tive and non-specific, the disease is usually diagnosed at later stages. Despite the availability of numerous possible treat- ments, hepatic metastases are difficult to eradicate because of their late discovery. Early and accurate detection of these lesions is recognized as having the potential of improving sur- vival rates and reducing treatment morbidity. A key observation is that there are changes in liver blood supply that can serve as an indicator for the presence of hep- atic metastases [1]. It is well known that, whereas normal liver is supplied predominantly by the portal vein, in patients with overt colorectal liver metastases, a higher proportion of liver blood flow is derived from the hepatic artery. Thus, by monitoring hemodynamical changes, earlier detection of hep- atic metastases may be feasible. Imaging plays a central role in the early diagnosis of liver metastases. The association between hepatic metastases and altered liver blood flow has been demonstrated recently by dynamic scintigraphy [1], by Doppler sonography [2, 3], by dynamic contrast enhanced CT [4] and more recently by dy- namic contrast enhanced MRI [5]. Measurements using MRI can potentially overcome limitations posed by other imag- ing techniques, such as poor spatial resolution in radionuclide studies, lack of reproducibility in Doppler US and radiation exposure using CT. Currently, acquisition of perfusion images in both CT and MRI require the intravenous administration of a contrast agent. Good separation of arterial from portal phase requires high temporal resolution which enforces reduction of the spatial resolution. In a previous work [6], we demonstrated the feasibil- ity of fMRI with hypercapnia and hyperoxia for monitoring changes in liver perfusion and hemodynamics without the need of a contrast agent administration. Using this method we characterized colorectal hepatic metastases and were able to follow their early hemodynamical changes [7, 8]. Since our method detects steady state levels without the use of contrast agent there is no need to compromise the spatial resolution and image quality. Therefore, we expect to be able to detect smaller lesions. However, the manual analysis of the hemo- dynamical maps produced by this method turned out to be a difficult, time consuming, and potentially unreliable task. In this work we present a machine learning approach for the automatic detection of colorectal hepatic metastases based on their hemodynamical changes. First, we construct a sta- tistical model describing the hemodynamical changes of col- orectal hepatic metastases from samples obtained at advanced phase of metastases growth, where the metastases were visi- ble in the anatomical MRI. Then, new samples obtained at the earlier phase of metastases growth, where the metastases are not yet visible in the anatomical image, are classified ac- 584 978-1-4244-2003-2/08/$25.00 ©2008 IEEE ISBI 2008