Interactive Framework for Insect Tracking with Active Learning
2014, Shen, Minmin, Huang, Wei, Szyszka, Paul, Galizia, C. Giovanni, Merhof, Dorit
Extracting motion trajectories of insects is an important prerequisite in many behavioral studies. Despite great efforts to design efficient automatic tracking algorithms, tracking errors are unavoidable. In this paper, we propose general principles that help to minimize the human effort required for accurate multi-target tracking in the form of applications that can track the antennae and mouthparts of a honey bee based on a set of low frame rate videos. This interactive framework estimates which key frames will require user correction, i.e. those that are used for user correction, which are used for 1) incrementally learning an object classifier and 2) data association based tracking. To this framework we apply a standard classification algorithm (i.e. naive Bayesian classification) and an association optimization algorithm (i.e. Hungarian algorithm). The precision of tracking results by our framework on real-world video data is above 98%.
Automated Image Processing to Quantify Cell Migration
2013-02-20, Shen, Minmin, Zimmer, Bastian, Leist, Marcel, Merhof, Dorit
Methods to evaluate migration capacity of stem cells and the inhibition by chemicals are important for biomedical research. Here, we established an automated image processing framework to quantify migration of human neural crest (NC) cells into an initially empty, circular region of interest (ROI). The ROI is partially filled during the experiment by migrating cells. Based on an image captured only once at the end of the biological experiment, the framework identifies the initial ROI. The identification worked also, when the distribution of surrounding cells showed large heterogeneity. After segmentation, the number of migrated cells was identified. The image processing framework was capable of efficiently quantifying chemical effects on cell migration.
Evaluation of Diffusion-Tensor Imaging based Global Search and Tractography for Tumor Surgery close to the Language System
2013, Richter, Mirco, Zolal, Amir, Ganslandt, Oliver, Buchfelder, Michael, Nimsky, Christopher, Merhof, Dorit
Pre-operative planning and intra-operative guidance in neurosurgery require detailed information about the location of functional areas and their anatomo-functional connectivity. In particular, regarding the language system, post-operative deficits such as aphasia can be avoided. By combining functional magnetic resonance imaging and diffusion tensor imaging, the connectivity between functional areas can be reconstructed by tractography techniques that need to cope with limitations such as limited resolution and low anisotropic diffusion close to functional areas. Tumors pose particular challenges because of edema, displacement effects on brain tissue and infiltration of white matter. Under these conditions, standard fiber tracking methods reconstruct pathways of insufficient quality. Therefore, robust global or probabilistic approaches are required. In this study, two commonly used standard fiber tracking algorithms, streamline propagation and tensor deflection, were compared with a previously published global search, Gibbs tracking and a connection-oriented probabilistic tractography approach. All methods were applied to reconstruct neuronal pathways of the language system of patients undergoing brain tumor surgery, and control subjects. Connections between Broca and Wernicke areas via the arcuate fasciculus (AF) and the inferior fronto-occipital fasciculus (IFOF) were validated by a clinical expert to ensure anatomical feasibility, and compared using distance- and diffusion-based similarity metrics to evaluate their agreement on pathway locations. For both patients and controls, a strong agreement between all methods was observed regarding the location of the AF. In case of the IFOF however, standard fiber tracking and Gibbs tracking predominantly identified the inferior longitudinal fasciculus that plays a secondary role in semantic language processing. In contrast, global search resolved connections in almost every case via the IFOF which could be confirmed by probabilistic fiber tracking. The results show that regarding the language system, our global search is superior to clinically applied conventional fiber tracking strategies with results similar to time-consuming global or probabilistic approaches.
INCIDE the Brain of a Bee : Visualising Honeybee Brain Activity in Real Time by Semantic Segmentation
2012-10, Strauch, Martin, Brög, Marc P., Müthing, Clemens, Szyska, Paul, Deussen, Oliver, Galizia, C. Giovanni, Merhof, Dorit
We present a software solution for processing recordings of honeybee brain activity in real time. In the honeybee brain, odors elicit spatio-temporal activity patterns that encode odor identity. These patterns of neural activity in units called glomeruli can be recorded by calcium imaging with fluorescent dyes, but so far glomerulus segmentation was only possible offline, making interactive experiments impossible. Our main contribution is an adaptive algorithm for image processing, along with a fast implementation for the graphics processing unit that enables semantic segmentation in real time. Semantics is based on the temporal dimension, relying on the fact that time series of pixels within a glomerulus are correlated. We evaluate our software on reference data, demonstrate applicability in a biological experiment, and provide free source code. This paves the way for interactive experiments where neural units can be selected online based on their past activity.
Evaluation of a human neurite growth assay as specific screen for developmental neurotoxicants
2013-12, Krug, Anne K., Stiegler, Nina, Matt, Florian, Schönenberger, Felix, Merhof, Dorit, Leist, Marcel
Organ-specific in vitro toxicity assays are often highly sensitive, but they lack specificity. We evaluated here examples of assay features that can affect test specificity, and some general procedures are suggested on how positive hits in complex biological assays may be defined. Differentiating human LUHMES cells were used as potential model for developmental neurotoxicity testing. Forty candidate toxicants were screened, and several hits were obtained and confirmed. Although the cells had a definitive neuronal phenotype, the use of a general cell death endpoint in these cultures did not allow specific identification of neurotoxicants. As alternative approach, neurite growth was measured as an organ-specific functional endpoint. We found that neurite extension of developing LUHMES was specifically inhibited by diverse compounds such as colchicine, vincristine, narciclasine, rotenone, cycloheximide, or diquat. These compounds reduced neurite growth at concentrations that did not compromise cell viability, and neurite growth was affected more potently than the integrity of developed neurites of mature neurons. A ratio of the EC50 values of neurite growth inhibition and cell death of >4 provided a robust classifier for compounds associated with a developmental neurotoxic hazard. Screening of unspecific toxicants in the test system always yielded ratios <4. The assay identified also compounds that accelerated neurite growth, such as the rho kinase pathway modifiers blebbistatin or thiazovivin. The negative effects of colchicine or rotenone were completely inhibited by a rho kinase inhibitor. In summary, we suggest that assays using functional endpoints (neurite growth) can specifically identify and characterize (developmental) neurotoxicants.
Optic radiation fiber tractography in glioma patients based on high angular resolution diffusion imaging with compressed sensing compared with diffusion tensor imaging : initial experience
2013, Kuhnt, Daniela, Bauer, Miriam H. A., Sommer, Jens, Merhof, Dorit, Nimsky, Christopher
Up to now, fiber tractography in the clinical routine is mostly based on diffusion tensor imaging (DTI). However, there are known drawbacks in the resolution of crossing or kissing fibers and in the vicinity of a tumor or edema. These restrictions can be overcome by tractography based on High Angular Resolution Diffusion Imaging (HARDI) which in turn requires larger numbers of gradients resulting in longer acquisition times. Using compressed sensing (CS) techniques, HARDI signals can be obtained by using less non-collinear diffusion gradients, thus enabling the use of HARDI-based fiber tractography in the clinical routine.
Eight patients with gliomas in the temporal lobe, in proximity to the optic radiation (OR), underwent 3T MRI including a diffusion-weighted dataset with 30 gradient directions. Fiber tractography of the OR using a deterministic streamline algorithm based on DTI was compared to tractography based on reconstructed diffusion signals using HARDI+CS.
HARDI+CS based tractography displayed the OR more conclusively compared to the DTI-based results in all eight cases. In particular, the potential of HARDI+CS-based tractography was observed for cases of high grade gliomas with significant peritumoral edema, larger tumor size or closer proximity of tumor and reconstructed fiber tract.
Overcoming the problem of long acquisition times, HARDI+CS seems to be a promising basis for fiber tractography of the OR in regions of disturbed diffusion, areas of high interest in glioma
Fiber tractography based on diffusion tensor imaging compared with high-angular-resolution diffusion imaging with compressed sensing : initial experience
2013-01, Kuhnt, Daniela, Bauer, Miriam H. A., Egger, Jan, Richter, Mirco, Kapur, Tina, Sommer, Jens, Merhof, Dorit, Nimsky, Christopher
The most frequently used method for fiber tractography based on diffusion tensor imaging (DTI) is associated with restrictions in the resolution of crossing or kissing fibers and in the vicinity of tumor or edema. Tractography based on high-angular-resolution diffusion imaging (HARDI) is capable of overcoming this restriction. With compressed sensing (CS) techniques, HARDI acquisitions with a smaller number of directional measurements can be used, thus enabling the use of HARDI-based fiber tractography in clinical practice.
To investigate whether HARDI+CS-based fiber tractography improves the display of neuroanatomically complex pathways and in areas of disturbed diffusion properties.
Six patients with gliomas in the vicinity of language-related areas underwent 3-T magnetic resonance imaging including a diffusion-weighted data set with 30 gradient directions. Additionally, functional magnetic resonance imaging for cortical language sites was obtained. Fiber tractography was performed with deterministic streamline algorithms based on DTI using 3 different software platforms. Additionally, tractography based on reconstructed diffusion signals using HARDI+CS was performed.
HARDI+CS-based tractography displayed more compact fiber bundles compared with the DTI-based results in all cases. In 3 cases, neuroanatomically plausible fiber bundles were displayed in the vicinity of tumor and peritumoral edema, which could not be traced on the basis of DTI. The curvature around the sylvian fissure was displayed properly in 6 cases and in only 2 cases with DTI-based tractography.
HARDI+CS seems to be a promising approach for fiber tractography in clinical practice for neuroanatomically complex fiber pathways and in areas of disturbed diffusion, overcoming the problem of long acquisition times.
Optimized cortical subdivision for classification of Alzheimer's disease with cortical thickness
2013-02-20, Richter, Mirco, Merhof, Dorit
In several studies, brain atrophy measured by cortical thickness has shown to be a meaningful biomarker for Alzheimer’s disease. In this research field, the level of granularity at which values are compared is an important aspect. Vertex- and voxel-based approaches can detect atrophy at a very fine scale, but are susceptible to noise from misregistrations and inter-subject differences in the population. Regional approaches are more robust to these kinds of noise, but cannot detect variances at a local scale. In this work, an optimized classifier is presented for a parcellation scheme that provides a trade-off between both paradigms by increasing the granularity of a regional approach. For this purpose, atlas regions are subdivided into gyral and sulcal parts at different height levels. Using two-stage feature selection, optimal gyral and sulcal subregions are determined for the final classification with sparse logistic regression. The robustness was assessed on clinical data by 10- fold cross-validation and by testing the prediction accuracy for unseen individuals. In every aspect, superior classification performance was observed as compared to the original parcellation scheme which can be
explained by the increased locality of cortical thickness measures and
the customized classification approach that reveals interacting regions.
A Combined MR-PET Analysis of Wholefield and Subfield Hippocampal Changes in AD and FTLD
2013, Bishop, Courtney A., Zamboni, Giovanna, Dukart, Jürgen, Müller, Karsten, Barthel, Henryk, Sabri, Osama, Schroeter, Matthias L., Declerck, Jerome, Merhof, Dorit, Jenkinson, Mark
Classification of neurodegenerative dementia by Gaussian Mixture Models applied to SPECT images
2012-10, Stühler, Elisabeth, Platsch, Günther, Weih, Markus, Kornhuber, Johannes, Kuwert, Torsten, Merhof, Dorit
Gaussian mixture (GM) models can be applied for statistical classification of various types of dementia. As opposed to linear boundaries, they do not only provide the class membership of a case, but also a measure of its probability. This enables an improved interpretation and classification of neurodegenerative dementia datasets which comprise various stages of the disease, and also mixed forms of dementia. In this work, GM models are applied to a total number of 103 technetium-99methylcysteinatedimer (99mTc-ECD) SPECT datasets of asymptomatic controls (CTR), as well as Alzheimer’s disease (AD) and frontotemporal dementia (FTD) patients in early or moderate stages of the disease. Prior to classification, multivariate analysis is applied: Principal component analysis (PCA) is used for dimensionality reduction, followed by a differentiation of the datasets via multiple discriminant analysis (MDA). A GM model on the resulting discrimination plane is constructed by computing the GM distribution associated with the underlying training set. The posterior probabilities of each case indicate its class membership probability. The performance of GM models for classification is assessed by bootstrap resampling and cross validation. Accuracy and robustness of the method are evaluated for different numbers of principal components (PCs), and furthermore the detection rate of dementia in early stages is calculated. The GM model outperfomes classification with linear boundaries in both predicted accuracy and detection rate of early dementia, and is equally robust.