· · · Repository hosted by TU Delft Library

In this paper, we propose some new features that are based on the correlation of deformation vectors of two or more brain structures when one MRI brain volume (the Template image) is transformed into another volume (the Study image) by a Computational Anatomy method. Thenew features can reveal the deformation relationship between two different structures in a quantitative way. We also use vector classification methods to distinguish Normal subjects (NL) from Alzheimer Disease (AD) patients using combined sets of known and new features ,and apply these methods to a set of MRI volumes from the ADNI database. Using these methods we obtain a good correspondence between theclassification outcomes and the ground truth data. Also, we visualize our data and results in a specially designed user interface. Conclusion:In this paper, we have defined a new set of parameters for the analysis of evolution of neurodegenerative diseases based on MR images. In order to be able to visualize the deformation and investigate the usefulness of these parameters, we developed a visualizationenvironment that displays and evaluates them, and helps to select the regions for which they should be calculated, and subsequently applied vector classification methods to monitor whether AD patients could be discerned from NL subjects. We have tested this on a data setfrom the ADNI database. Our results show that these parameters do indeed indicate differences between AD and NL subjects. From the cross validation result, we find that the Directional Correlation Coefficient of the ventricle and hippocampus, Mean Jacobian Displacement Correlation of the ventricle and hippocampus and Ventricle Region size are three key features that are most promising as parameters todistinguish AD patients from NL subjects. Of the three considered methods, the SVM classification method is the best method to make anautomatic classification for this task.

Mood is an important aspect of music and knowledge on mood can be used as a basic ingredient in music recommender and retrieval systems.A music experiment was carried out establishing ratings for variousmoods and a number of attributes like valence and arousal. The analysis of these data is presented in this paper covering the issues ofthe number of basic dimensions in music mood, their relation to valence and arousal, the distribution of moods in the valence-arousal plane, distinctiveness of the labels, and appropriate (number of) labels for full coverage of the plane. It is also shown that subject-averaged valence and arousal ratings can be predicted from music features by a linear model.

We report on the use of diffuse optical spectroscopy analysis of breast spectra acquired in the wavelength range from 500 to 1600 nm with a fiber optic probe. A total of 102 ex vivo samples of five different breast tissue types, namely adipose, glandular, fibroadenoma, invasive carcinoma and ductal carcinoma in situ from 52 patients weremeasured. A model deriving from the diffusion theory was applied tothe measured spectra in order to extract clinically relevant parameters such as blood, water, lipid, and collagen volume fractions, b-carotene concentration, average vessels radius, reduced scattering amplitude, Mie slope and Mie-to-total scattering fraction. Based on a classification and regression tree algorithm applied to the derived parameters, a sensitivity-specificity of 98%-99%, 84%-95%, 81%-98%, 91%-95%, and 83%-99% were obtained for discrimination of adipose, glandular, fibroadenoma, invasive carcinoma, and ductal carcinomain situ, respectively; and a multiple classes overall diagnostic performance of 94%. Sensitivity-specificity values obtained for discriminating malignant from non-malignant tissue were compared to existing reported studies by applying the different classification methodsthat were used in each of these studies. Furthermore, in these reported studies, either lipid or b-carotene was considered as adipose tissue precursors. We estimate both chromophore concentrations and demonstrate that lipid is a better discriminator for adipose tissue than b-carotene.

Dementia is significant loss of intellectual abilities such as memory capacity, severe enough to interfere with social or occupational functioning. The most common types of dementia are: Alzheimer's disease (AD), Lewy body dementia (LBD), and frontotemporal dementia (FTD). In 2008, there are currently 29.8 million people with dementia, with the number expected to be 81.1 million by 2050. The classification of FDG-PET in patients with dementia might not be an easy task. Some dementia diseases have similar disease patterns which lead to the misdiagnosis of these diseases. Nowadays, images are visually evaluated by an expert reader and this process is not entirely quantitative or reproducible. Even the experts can have up to 20% misclassification. Automated diagnosis by pattern recognition can produce quantitative and reproducible results, but if training data comes from clinical routine, it may produce less accurate results to discriminate similar disease patterns. Hence, the work of this thesis aims to find an optimal subset of features, for a given training data set, which improves the classification of FDG-PET in patients with suspected dementia.The presented work describes some methods that found effective in improving the classification problem at hand. These methods consist of: (i) feature extraction, (ii) feature ranking, (iii) classifier learning with balanced data, (iv) feature selection. The results demonstrate the possibility to improve the accuracy of pair-wise classification for the most common dementia diseases with an excellent accuracy(in less time) by these methods.

We investigate the potential to enhance breast cancer event predictors by exploiting subtype information. We do this with a two-stage approach that first determines a sample's subtype using a recent module-driven approach, and secondly constructs a subtype-specific predictor to predict a metastasis event within five years. Our methodology is validated on a large compendium of microarray breast cancer datasets,including 43 replicate array pairs for assessing subtyping stability. Note that stratifying by subtype strongly reduces the training set sizes available to construct the individual predictors, which may decrease performance. Besides sample size, other factors likeunequal class distributions and differences in the number of samplesper subtype, easily obscure a fair comparison between subtype-specific predictors constructed on different subtypes, but also between subtype specific and subtype a-specific predictors. Therefore, we constructed a completely balanced experimental design, in which none ofthe above factors play a role and show that subtype-specific eventpredictors clearly outperform predictors that do not take subtype information into account.

Recommender systems are becoming popular tools to aid users in finding interesting and relevant TV-shows and other digital video assets,based on implicitly learned user preferences. In this context, a common assumption is that user preferences can be specified by program types (movie, sports, ...) and that an asset can be labeled by oneor more program types, thus allowing an initial coarse pre-selection of potentially interesting assets. Furthermore each asset has a short textual description, which allows us to investigate whether it might be useful to automatically label assets with program type labels. To that purpose we compared the Vector Space Model with more recent approaches to text classification, such as Logistic Regression and Random Indexing on a large collection of TV-shows descriptions. The experimental results show that LR is the best approach, but RI outperforms VSM under particular conditions.

Motivation: In recent years increasing evidence appeared that breastcancer may not constitute a single disease at the molecular level,but comprises a heterogeneous set of subtypes. This suggests that instead of building a single predictor, better predictors might be constructed that solely target samples of a designated subtype. An unavoidable drawback of developing subtype-specific predictors, however,is that a stratification by subtype drastically reduces the numberof samples available for their construction. It is therefore questionable whether the potential benefit of subtyping can outweigh the drawback of a severe loss in sample size. Factors like unequal class distributions and differences in the number of samples per subtype, further complicate comparisons. Results: We present several evaluation strategies that facilitate a comprehensive comparison between subtype-specific predictors and predictors that do not take subtype information into account. Emphasis lies on careful control of sample size as well as class and subtype distributions. The methodology is applied to a large breast cancer compendium involving over 1500 arrays,using a state-of-the-art subtyping scheme. We show that the resulting subtype-specific predictors outperform those that do not take subtype information into account, especially when taking sample size considerations into account.

Mood is an important attribute of music and knowledge on mood can beused as a basic ingredient in music recommender and retrieval systems. Moods are assumed to be dominantly determined by two dimensions:valence and arousal. An experiment was conducted to attain data forsong-based ratings of valence and arousal. It is shown that the subject-averaged ratings can be predicted from music features by a linear model.

Medical Image Classification& Retrieval is nowadays a hot topic. Many projects already exist in this field, like the IRMA project, but the obtained results are not yet good enough for real implementation. The goal of this report is to investigate in features that could enable a good performance for a multiclass classification task. First the database used for this student project is presented. Then, several features are described and evaluated on a benchmarked dataset. Finally, after conclusion, some indications of future work to improve the performance of such system are given.

For many speech telecommunication technologies a robust speech activity detector is important. An audio-only speech detector will givefalse positives when the interfering signal is speech or has speech characteristics. The modality video is suitable to solve this problem. In this report the approach to and implementation of a decision-based audiovisual speech detector is given. Acoustic and visual features of speech are first separately investigated. Firstly, a common method for speech detection based on audio has been built. Secondly, from the video data the mouth features have been extracted with the implementation of an own idea. The visual features were used to create a conservative visual non-speech detector. The low false detection rate makes the visual non-speech detector suitable to rule out some false speech detections of an audio only solution. Finally, the combinationof the audio detector and the video detector leads to an audiovisualspeech detector which uses basic mouth features and a common acousti-cal speech detection method to outperform an audio-only solution.

The aging population in the European Union and the US has increased the importance of research in neurodegenerative diseases. Imaging plays an essential role in this endeavor by providing insight to the intricate cellular and inter-cellular processes in living tissues that will otherwise be difficult, or impossible, to gain. Because of the sheer size of the imagery data, the lack of sufficient medical staff, and the inaccuracies resulting from manual processing, automated processing of image-based data to generate quantitative and reproducible results is necessary. To this effect, in this thesis a fully automatic image-processing algorithm for brain tissue segmentation from magnetic resonance (MR) images is proposed. Contrary to the present iterative expectation maximization (EM) based algorithms, it uses online (sample-by-sample) learning to adapt to the intensity inhomogeneity inherent to MR images. Since the proposed method can adapt to the intensity inhomogeneity online, multiple iterations over the data as in the present algorithms are not necessary, and consequently the processing time is decreased dramatically. The used online learning scheme is based on learning vector quantization and is further tailored to the segmentation of MR images by integration of spatial context and the use of a special locality-preserving scanning order of the data. Explorations of various scanning orders and a modification to the learning rule to allow for 3D learning have lead to three variants of the proposed algorithm. These proposed methods are validated by comparing the segmentation masks to basic k-means clustering, and present EM-based methods, namely, FAST and the state-of-the-art EMS, on simulated and real datasets. The proposed methods demonstrated a significant reduction of the processing time (a factor of 20) compared to the EM-based methods. Tests on BrainWeb simulated data showed that segmentation accuracy is comparable to the EM-based methods, however, tests on real data, where the segmentations of EMS were used as ground truth, showed lower accuracy than the EM-based FAST. Moreover, the tests on real data showed that the proposed methods as well as FAST make a significant amount of misclassifications in the so-called deep grey matter areas, which suggests the necessity of a spatial prior atlas as it is used in EMS.

Environmental sound classification has a wide range of applications,like hearing aids, mobile communication devices, portable media players, and auditory protection devices. Sound classification systemstypically extract features from the input sound. Using too many features increases complexity unnecessarily and can even reduce performance, but proper selection of features is a non-trivial task. It is promising to base such selection on soft computing, because this approach can handle uncertain information in an efficient way, using simple set theoretic functions, and because this approach is close to perception-based reasoning. Therefore, this thesis investigates different feature selection methods, including soft computing methods andclassical information, entropy and correlation-based approaches. Results show that a soft computing based feature selection method performs best in terms of number of features selected, recognition rateand consistency of performance. In addition, the resulting classification system is robust for reverberation.