Enhancing Pancreatic Adenocarcinoma Delineation in Diffusion Derived Intravoxel Incoherent Motion f-Maps Through Automatic Vessel and Duct Segmentation Thomas J. Re, 1,2 Andreas Lemke, 3 Miriam Klauss, 4 Fredrik B. Laun, 5 Dirk Simon, 1 Katharina Gru ¨ nberg, 1 Stefan Delorme, 1 Lars Grenacher, 3 Riccardo Manfredi, 2 Roberto Pozzi Mucelli, 2 and Bram Stieltjes 6 * Diffusion-based intravoxel incoherent motion imaging has recently gained interest as a method to detect and character- ize pancreatic lesions, especially as it could provide a radia- tion- and contrast agent-free alternative to existing diagnostic methods. However, tumor delineation on intravoxel incoherent motion-derived parameter maps is impeded by poor lesion-to- pancreatic duct contrast in the f-maps and poor lesion-to- vessel contrast in the D-maps. The distribution of the diffusion and perfusion parameters within vessels, ducts, and tumors were extracted from a group of 42 patients with pancreatic ad- enocarcinoma. Clearly separable combinations of f and D were observed, and receiver operating characteristic analysis was used to determine the optimal cutoff values for an automated segmentation of vessels and ducts to improve lesion detection and delineation on the individual intravoxel incoherent motion- derived maps. Receiver operating characteristic analysis identi- fied f 5 0.28 as the cutoff for vessels (Area under the curve (AUC) 5 0.901) versus tumor/duct and D 5 1.85 mm 2 /ms for sep- arating duct from tumor tissue (AUC 5 0.988). These values were incorporated in an automatic segmentation algorithm and then applied to 42 patients. This yielded clearly improved tumor delineation compared to individual intravoxel incoherent motion- derived maps. Furthermore, previous findings that indicated that the f value in pancreatic cancer is strongly reduced compared to healthy pancreatic tissue were reconfirmed. Magn Reson Med 66:1327–1332, 2011. V C 2011 Wiley Periodicals, Inc. Key words: IVIM; segmentation; pancreas; diffusion Pancreatic adenocarcinoma is a severe disease with poor prognosis; currently, early radical surgery is providing the only hope of survival (1,2). Radiological detection and characterization of this disease must therefore offer high sensitivity for early detection, high specificity to prevent unnecessary surgery, and delineation accuracy for appropriate surgical planning (3). Current standard imaging of pancreatic lesions often uses a combination of ultrasound, MR, and computed tomography (CT) to increase diagnostic certainty (4–6). Emerging pancreatic lesion imaging modalities, such as those based on diffu- sion-weighted (DW) MR, and in particular diffusion- derived perfusion imaging could prove useful in improv- ing early detection and diagnostic certainty if considered in conjunction with these existing techniques. Technical advances in the last decade have enabled the use of DW imaging (DWI) of the abdomen where organ motion and tissue inhomogeneity have long pre- vented robust diffusion measurements (7–9). In particu- lar, some reports have shown that the DWI-derived parameter apparent diffusion coefficient (ADC) can be used to detect and characterize cystic and cancerous lesions of the pancreas as well as other abdominal organs; pancreatic adenocarcinoma tissue has a significantly lower ADC value than that of normal pancreatic tissue (8–13). More recent studies (14,15) have investigated the use of other DWI-derived parameters, specifically those based on the intravoxel incoherent motion (IVIM) model first pre- sented by Le Bihan in 1986 (16–19). In this model, the DW signal is considered to be influenced by two distinct types of molecular motion: diffusion, due to thermal Brownian motion and perfusion, due to the microcirculation in capil- laries. The IVIM model is represented mathematically by the equation first described by Le Bihan (17) S S 0 ¼ð1  f Þ expðbDÞþ f  expðb ðD þ D  ÞÞ: ½1 Here, S/S 0 is the signal decay measured in each indi- vidual voxel in the DWI, the b value represents the strength of the diffusion weighting, D is the diffusion constant, D* is the pseudo-diffusion coefficient, and f is the perfusion fraction. This equation is biexponential with the first term representing primarily the influence of diffusion in tissue on the DW signal, while the second term represents the vascular component contribution mainly due to perfusion. In a recent work, the IVIM perfusion fraction f was shown to be significantly lower in pancreatic tumors than in healthy pancreatic tissue and to be quantitatively superior to both ADC and IVIM diffusion coefficient D for characterizing pancreatic lesions (20). Furthermore, 1 Department of Radiology, German Cancer Research Center, Heidelberg, Germany. 2 Department of Radiology, University of Verona, Verona, Italy. 3 Department of Diagnostic Radiology, University of Heidelberg, Heidelberg, Germany. 4 Computer Assisted Clinical Medicine, University of Heidelberg, Faculty of Medicine, Mannheim, Germany. 5 Department of Medical Physics in Radiology, German Cancer Research Center, Heidelberg, Germany. 6 Quantitative Imaging Based Disease Characterization, German Cancer Research Center, Heidelberg, Germany. *Correspondence to: Bram Stieltjes, MD, DKFZ Section, Quantitative Imaging Based Disease Characterization (E011), INF 280, 69120 Heidelberg, Germany. E-mail: b.stieltjes@dkfz.de Received 7 May 2010; revised 25 February 2011; accepted 26 February 2011. DOI 10.1002/mrm.22931 Published online 24 March 2011 in Wiley Online Library (wileyonlinelibrary. com). Magnetic Resonance in Medicine 66:1327–1332 (2011) V C 2011 Wiley Periodicals, Inc. 1327