شناسایی اتوماتیک مقادیر گمراه کننده جریان خون در مطالعات سی تی اسکن پرفیوژن سرطان ریه  Automatic detection of misleading blood flow values in CT perfusion studies of lung cancer

Abstract In the oncology field, the anti-angiogenetic therapies aim at inhibiting tumour vascularization, that is the development of new capillary blood vessels in tumours, that allows them to grow and spread and, potentially, to metastasi. Computed tomography perfusion (CTp) is a dynamic contrast-enhanced technique that has emerged in the last few years as a promising approach for earlier assessment of such therapies, and of tumour response, in general, since functional changes precede morphological changes, that take more time to become evident. However several issues, such as patient motion and several types of artefacts, jeopardize quantitative measurements,this preventing CTp to be used in standard clinics. This paper presents an original automatic approach, based on the voxel-based analysis of the time–concentration curves (TCCs), that allows emphasizing those physiological structures, such as vessels, bronchi or artefacts, that could affect the final computation of blood flow perfusion values in CTp studies of lung cancer. The automatic exclusion of these misleading values represents a step towards a quantitative CTp, hence its routine use in clinics. ۱٫ Introduction In the last years the computed tomography perfusion (CTp) has aroused lively interest because of its capability of providing morphologically detailed functional maps that could find application in monitoring functional activity of tumours at their different stages [1], predicting treatment outcome or early therapeutic response of anti-angiogenetic therapies [2–۶], before morphological changes become visible [7]. This widely available and non-invasive technique relies on the estimation of tissue contrast agent delivery, and on the corresponding haemodynamic parameters, which can be derived from the analysis of the tissue time-concentration curves (TCCs) signals generated by the contrast agent before, during, and after reaching the tumour lesion. As such, they can be employed to detect changes in its vascular structure, hinting at possible anomalies in blood supply (i.e., tumour angiogenesis [2,8]). The main obstacles preventing the use of CTp in the standard clinical practice is represented by the difficulty to measure its reproducibility and  in the last analysis, the reliability of perfusion values [9]. In its turn, this arises from the difficulty to achieve reliable TCCs because of motion artefacts due to breathing [10] or a high concentration of contrast agent that may also yield streaking [11] or simulate partial volume effects in the repeated acquisitions of a CTp study. In addition, streaks or dark bands may also originate from beam hardening (for instance, induced by the presence in the lesion’s neighbourhood of a main vessel flooded with contrast agent). A detailed description of possible artefacts in CT examinations, also holding for CTp examinations, can be found in [12,13]. Besides artefacts [14], also vessels [15] and bronchi [16] are usually manually excluded from the perfusion analysis not to jeopardize the visual analysis of perfusion colour maps nor the automatic computation of local or global statistical indexes regarding perfusion values. In general, the unreliable perfusion values in colour maps are simply considered as those being out of range of physiological parameters (for instance, due to vessels [17] and are detected and excluded by manually adjusting an appropriate window level [9]. So far, the TCC fitting errors and goodness of fit indexes have been mainly used to evaluate the reliability of given simulated model fitting, from a theoretical point of view, in lung CT [18] or liver MRI [19] perfusion studies, rather than to assess the voxel-based reliability of perfusion values. Just recently, some works have appeared to try addressing explicitly the reliability of TCCs in relation to the outcome of perfusion studies, suggesting methods to detect where the error in fitting a TCC according to a given pharmacokinetic model could prejudice the computation of correct perfusion values [20] or even to improve the way a TCC is built, so as to reduce the number of possible flawed TCCs [21]. However, they do not associate fitting errors to their causes, if not to generic motion artefacts and, not at all, to anatomical structures. This methodological work presents a novel quantitative and automatic approach to detect those anatomical structures (mainly vessels and bronchi) and those regions undergoing CTp reconstruction and acquisition artefacts, that could compromise the correct interpretation of a CTp colour map and, ultimately, the clinical outcome. The approach is based on the computation of a statistical error index connected to the quality (i.e., goodness of fit) of TCCs. The ability of our method to automatically remove the “misleading” regions is assessed and compared with the performance of two 25-year experienced radiologists who detected, and manually bounded using a graphic tablet, the anatomical structures and the regions undergoing artefacts. The main errors made when operating manually, and their consequences on the definite perfusion maps, are also discussed. Moreover, we analyze how mean of perfusion values and, above, their standard deviation and coefficient of variation (CV), change before and after removing automatically segmented regions. Finally, we also discuss some meaningful comparisons between definite colour maps achieved by using our approach and the manual thresholding on BF values commonly used by readers. The paper is organized as follows. Section 2 describes the materials and the methods employed from image acquisition to building of the final perfusion maps, including the automatic histogrambased image segmentation approach developed by exploiting goodness-of-fit errors. Section 3 presents the choices performed to evaluate the experimental results which are subsequently analysed in Section 4 and discussed in Section 5. Finally, drawing conclusions are reported in Section 6.