• Home
• Preliminary Program
• Keynote speakers
• Registration & Venue
• Company Market/sponsoring
• Important Dates
• Awards
• Organizing Board
• Conference Topics
• Call for Abstracts
• Contact

10:30   Biomedical Image Processing Chair: Theo van Walsum 10:30 15 mins INTER-OBSERVER VARIATION IN SEGMENTING GLIOMA ON MRI BEFORE AND AFTER TREATMENT Martin Visser, Domenique Muller, Roelant Eijgelaar, Jan de Munck, Philip de Witt Hamer, Marnix Witte, Frederik Barkhof, Niels Verburg Abstract: Quantitation of glioma volume and location on MRI is becoming standard procedure in the evaluation of treatment and follow-up. Manual delineation of glioma by an expert neuro-radiologist is considered to be the gold standard. Prior studies1 have demonstrated that different raters are in reasonable agreement when delineating the enhancing part of glioblastoma before treatment. Reproducible and reliable measurement of volume and location is necessary to obtain valid results, giving rise to the need to extend previous work. In this study the agreement between different raters is recorded in both enhancing and non-enhancing glioma before and after surgical treatment, and at progression (pre-op, post-op and progression respectively). A total of 40 patients with glioma were included, of which 20 are non-enhancing glioma, usually of WHO grade II or III, and 20 are enhancing, usually glioblastoma. Enhancing and non-enhancing elements were segmented by each rater for every patient at the three time points using dedicated Brainlab software made available as an online application. Inter-rater agreement is recorded with the generalized conformity index (CIgen)2, which is a generalization of the Jaccard score for two or more observers. Three raters, employed at the neurosurgery department of VUmc, have performed the segmentations. Patients with non-enhancing glioma show no enhancing elements with the exception for 3 patients in the progression time point and 1 patient in pre-op. Segmenters do not agree on the presence of residual tumor for all post-op moments. The segmented non-enhancing elements have a median CIgen score of 0.64, 0.30, and 0.27 in the pre-op, post-op and progression respectively. For patients with enhancing glioma both enhancing and non-enhancing elements are observed. In not all post-op moments segmenters agree on the presence of residual tumor. The segmented enhancing elements have a median CIgen score of 0.86, 0.24, and 0.69 in pre-op, post-op and progression respectively. Median CIgen for the non-enhancing elements is 0.47, 0.07 and 0.20 in pre-op, post-op and progression respectively. For segmentations in both the non-enhancing and enhancing glioma the agreement is highest when segmenting pre-op and lowest when segmenting post-op. The agreement between raters is higher for the enhancing elements than that for the non-enhancing elements in each respective time point for patients with enhancing glioma. Segmenter agreement is higher for all time points when segmenting enhancing glioma than when segmenting non-enhancing glioma. 10:45 15 mins 3D REGISTRATION OF MRI, HISTOLOGY AND COMPUTATIONAL FLUID DYNAMICS DATA OF ATHEROSCLEROTIC ARTERIES Astrid Moerman, Erik Jan Postema, Kristine Dilba, Frank Gijsen, Aad van der Lugt, Dirk Poot, Stefan Klein, Anton van der Steen, Kim van der Heiden Abstract: Atherosclerosis is characterized by the accumulation of lipid and inflammatory cells in the arterial wall, resulting in plaque formation. Although risk factors for atherosclerosis are systemic in nature, plaques develop at specific sites in the arterial tree. At these predilection sites the blood-exerted wall shear stress is low, resulting in activation of pro-inflammatory pathways in endothelial cells, which initiates plaque formation. At more advanced stages of the disease, the plaque intrudes into the lumen, affecting local shear stress patterns. Shear stress increases on the upstream side and throat of the lumen-intruding plaque, while shear stress remains low on the downstream side. Strikingly, rupture of vulnerable plaques occurs most frequently at high shear stress-exposed upstream sides. We hypothesize that alterations in the shear stress patterns during plaque growth influence plaque composition and play a role in plaque destabilization. To investigate the effect of shear stress on atherosclerotic plaque composition, we developed a pipeline to register 3D shear stress maps to 2D histological cross sections of human carotid plaques. As input parameters for the 3D shear stress map we obtained in vivo lumen geometry by MR imaging using a black blood stabilized 3D FSE sequence, as well as patient-specific time-dependent in- and outlet flow profiles, measured by 3D PC-MRI. A mesh was generated after manual lumen segmentation. CFD simulations were performed according to standard numerical procedures and 3D wall shear stress maps were generated. Of that same patient, the carotid plaque was surgically removed after MRI. The excised plaque was cut into 1 mm segments, which were each histologically processed and cut into 5 µm sections. We visualised plaque components using histochemical staining procedures (haematoxylin-eosin and resorcin-fuchsin) and also identified markers of plaque vulnerability, i.e. cap thickness, necrotic core size and macrophage infiltration. To enable the registration of each histology section we included additional imaging steps in our procedure, i.e. ex vivo MRI after surgical excision and photographs before each histological processing step. Rigid and/or affine registration steps were executed to deform and relocate the 2D histology sections to the 3D shear stress map. We successfully applied this registration pipeline on carotid endarterectomy samples of two patients. We plan to apply this method to a larger dataset in order to investigate the correlation between shear stress and plaque composition. This pipeline can also be of use in the validation of image segmentation algorithms. 11:00 15 mins CLASSIFYING GBM FOLLOW-UP OUTCOME USING SEMI-MANUAL DELINEATIONS AND MULTI-PARAMETRIC MRI Adrian Ion-Margineanu, Sofie Van Cauter, Diana Sima, Frederik Maes, Stefan Sunaert, Uwe Himmelreich, Sabine Van Huffel Abstract: Purpose. Delineating contrast enhancing (CE) tissue on T1 post-contrast Magnetic Resonance Images (T1pc MRI) is a key part of the Response Assessment in Neuro-Oncology (RANO) criteria for therapy follow-up in high-grade gliomas. The focus of this study is two-fold: (1) to evaluate the impact of semi-automatic delineation of hotspots of CE (HCE) in brain tumour follow-up of glioblastoma multiforme (GBM) patients after surgery, and (2) to evaluate results obtained by conventional and advanced MRI. Methods. Twenty-nine GBM patients who underwent surgery were scanned using a 3T MRI unit (Philips Achieva, Best, Netherlands). The protocol consisted of conventional MRI (cMRI: T1pc and T2), diffusion kurtosis imaging (DKI), and perfusion-weighted MRI (PWI: dynamic-susceptibility weighted contrast MRI). Regions of interest (ROIs) were manually drawn on T1pc images by an expert radiologist around the solid contrast-enhancing (CE) and Total tumour [1]. Another ROI was drawn around the contralateral normal appearing white matter to standardize MRI measurements. A label (responsive or progressive) was put on each patient according to the RANO criteria. We create another set of ROIs in the following way: for each session of each patient we use the T1pc intensities of all manually delineated voxels (Total ROI) and set an intensity threshold at the 90% percentile (P90). Voxels having intensities higher than or equal to P90 belong to HCE. After quantifying advanced MRI we obtain Cerebral Blood Volume and Cerebral Blood Flow from PWI, and Fractional Anisotropy and Mean Diffusivity from DKI. To these we add the two conventional maps, T1pc and T2, summing up to 6 parameter maps. For each time point of each patient, we perform an affine coregistration of the parameter maps to the T1pc map. For each parameter map and for both CE and HCE we compute the average and the 90th percentile of voxel intensities. In total, from 29 patients, we have 43 data points with manual CE ROIs, each with 12 features. After using our method to impute features, we obtain a 55 points dataset. Results. We tune Support Vector Machines with Gaussian kernel on separate feature subsets and report the balanced accuracy rate (BAR). Using CE features on the 43 points dataset the maximum BAR is 76% for PWI or DKI subset. Using HCE features on the same 43 points dataset the maximum BAR is 84% for cMRI. Using HCE features on the extended 55 points dataset the maximum BAR is 89% for PWI, followed by 85% for cMRI. Discussion. Semi-automatic delineations can improve classification of progressive vs. responsive patients. Moreover, data imputation using our method allows for more accurate classification. REFERENCES [1] S. Van Cauter et al., “Integrating diffusion kurtosis imaging, dynamic susceptibility-weighted contrast enhanced MRI, and short echo time chemical shift imaging for grading gliomas”, Neuro-Oncology, vol. 16, no. 7, pp. 1010–1021, 2014. 11:15 15 mins ACCURACY ESTIMATION FOR MEDICAL IMAGE REGISTRATION USING REGRESSION FORESTS Hessam Sokooti, Gorkem Saygili, Ben Glocker, Boudewijn Lelieveldt, Marius Staring Abstract: Most image registration methods do not provide insights about the quality of their results and devolve this difficult task to human experts, which is very time-consuming. Automatic evaluation of registration reduces the time of manual assessment and can provide information about the registration uncertainty. This is useful for refining the registration, either automatically or with the feedback of human experts. Even if refinement is not possible, information about the registration quality can help decide if subsequent processing is meaningful. Visualizing the error can be helpful in medical applications before making a clinical decision. This paper reports a new automatic algorithm to estimate mis-registration in a quantitative manner. A random regression forest is constructed using local and modality independent features related to the registration precision, the transformation model and intensity-based dissimilarity after registration. We introduce several new registration-based features based on the idea that the initial parameters of an optimization problem can affect the final solution for many registration paradigms. The variation in the final transformation result is then an intuitive measure for the local registration uncertainty, which is a surrogate for the correctness or at least the precision of the registration. In total, 7 mother features are used and the whole set of features is increased by generating a pool from each mother feature by calculating local averages and maxima using differently sized boxes. The method was evaluated on the SPREAD database [1], which contains 21 pairs of 3D lung CT images. For each patient a baseline and follow-up image (after 30 months) are available in which 100 well-distributed corresponding landmarks are selected semi-automatically on distinctive locations. The residual Euclidean distance after registration between the corresponding points can be seen as the accuracy of the registration, and is used to train and evaluate the forest. The results show that the mean absolute error of regression is 0.72 ± 0.96 mm and the accuracy of classification in three classes (correct, poor and wrong registration) is 93.4% with a precision of 95.8%, 38.9% and 69.7% in each class, respectively, comparing favorably to a competing method. In conclusion, the proposed method is for the first time able to automatically assess registration uncertainty by means of regression. Moreover, all features are modality-independent and the method can be used in combination with any registration paradigms, both for group-wise and pair-wise registrations. REFERENCES [1] J. Stolk, H. Putter, E. M. Bakker, S. B. Shaker, D. G. Parr, E. Piitulainen, E. W. Russi, E. Grebski, A. Dirksen, R. A. Stockely, J. H. Reiber and B. C. Stoel, “Progression parameters for emphysema: A clinical investigation,” Respiratory medicine, pp. 1924-1930, 2007. 11:30 15 mins VIDEO-BASED MONITORING AND ASSESSING DISCOMFORT IN PREMATURE INFANTS Yue Sun, Sveta Zinger, Sidarto Bambang Oetomo, Peter de With Abstract: Background. Premature infants are particularly vulnerable to the effects of pain and discomfort, which could lead to abnormal brain development, yielding long-term adverse neurodevelopmental outcomes [1]. They receive special care in the Neonatal Intensive Care ( NICU), where their vital signs are continuously monitored. However, there is currently no system for monitoring and detecting their pain or discomfort. We propose to employ video signals in order to detect infants’ discomfort automatically by analysing their facial expressions. Methods. We start by extracting faces of preterm infants on a frame-by-frame basis by applying Viola Jones face detection algorithm [2]. Following face detection, the extracted faces are classified for comfort or discomfort using a support vector machine classifier (SVM). The classifier is trained on four features, which are all selected according to primal face of pain. The eyes’ and mouth’s areas are calculated as two features by thresholding greyscale images of infants’ faces since squeezed eyes and stretched mouth are the signs pointing at discomfort. We also use the median of directional gradient along horizontal direction of infant’s face, which aims to capture the gradient information of nasolabial furrow and brow bulge. Moreover, the median of mutual information between 100 reference data of comfort faces and each training/testing frame is measured as a feature to indicate the likelihood between the two images. Results. Classification experiments are performed on 8 videos from 8 infants recorded at the Maxima Medical Center (MMC) in Veldhoven. The results show an overall AUC of 0.81 when all the features are combined. When we only use eyes’ area, mouth’s area, directional gradient and mutual information, we obtain the AUCs of 0.69, 0.57, 0.64, and 0.69, respectively. Conclusions. We perform classification on videos of preterm infants in order to detect their discomfort by analysing the facial expressions using 4 features. The result of an AUC of 0.81 shows the efficiency when combining the features. For future work, the system can be improved by adding more input features and achieving a higher scoring for previously unseen babies. REFERENCES [1] R. Whit Hall, and K. J. S. Anand. "Short-and long-term impact of neonatal pain and stress." Neoreviews 6.2 (2005): e69-e75. [2] P. Viola, and M. J. Jones. "Robust real-time face detection." International journal of computer vision 57.2 (2004): 137-154. 11:45 15 mins CSF CONTAMINATION-INVARIANT STATISTICS IN DIFFUSION-WEIGHTED MRI OF THE FORNIX Joor Arkesteijn, Dirk Poot, Marius de Groot, Arfan Ikram, Meike Vernooij, Wiro Niessen, Lucas van Vliet, Frans Vos Abstract: Introduction: Diffusion-weighted MRI is a frequently used technique to study the brain’s white matter. Diffusion measurements in white matter structures bordering cerebrospinal fluid (CSF) are susceptible to partial volume effects with CSF. The increased isotropic diffusion in CSF contaminated white matter voxels results in an overestimation of mean diffusivity (MD) and an underestimation of fractional anisotropy (FA). The fornix, the primary white matter bundle connecting the hippocampus to the mammillary bodies of the hypothalamus, is particularly prone to CSF contamination due to its small size and proximity to the third and lateral ventricles [1]. In this abstract we propose a method to discriminate between macrostructural and microstructural change in the body of the fornix based on conventional diffusion-weighted MRI data. Methods: A bi-tensor (BT) model [2], having an isotropic and an anisotropic diffusion compartment was fitted to the diffusion-weighted images (DWIs). Two approaches were used to make estimation feasible with DWIs acquired at a single (non-zero) b-value, namely using a global and a subject-specific trace-constraint of the diffusion tensor modelling the anisotropic compartment. The bi-tensor fractional anisotropy (BT-FA) and tissue fraction (BT-f) were analysed and compared to conventional diffusion statistics in simulated fiber bundles as well as in automated fornix segmentations of 577 elderly subjects from the Rotterdam Study. Linear regression was used to study the relation between diffusion statistics and age, with and without correcting for the cross-sectional area of the fornix. Results: In simulated fiber bundles the BT-FA was, unlike conventional diffusion statistics such as single-tensor fractional anisotropy (ST-FA) or mean diffusivity (ST-MD), unaffected by fiber bundle diameter. Cross-sectional area of the fornix decreased significantly with age. BT-FA in the population-based study did not correlate with age, whereas ST-FA, ST-MD and BT-f correlated significantly with age, even when correcting for the cross-sectional area of the fornix. Conclusion: Our findings suggest that age-related change to the body of the fornix consists of both macrostructural and microstructural effects. The distinction of an isotropic and an anisotropic diffusion compartment may allow a more sophisticated analysis in future studies of the fornix.