Szidónia Lefkovits

Brain Tumor Segmentation with Optimized Random Forest

In this paper we propose and tune a discriminative model based on Random Forest (RF) to accomplish brain tumor segmentation in multimodal MR images. The objective of tuning is meant to establish the optimal parameter values and the most significant constraints of the discriminative model. During the building of the RF classifier, the algorithm evaluates the importance of variables, the proximities between data instances and the generalized error. These three properties of RF are employed to optimize the segmentation framework. At the beginning the RF is tuned for variable importance evaluation, and after that it is used to optimize the segmentation framework. The framework was tested on unseen test images from BRATS. The results obtained are similar to the best ones presented in previous BRATS Challenges.

Automatic Brain Tumor Segmentation in Multispectral MRI Volumes Using a Random Forest Approach

The development of automatic tumor detection and segmentation procedures enables the computers to preprocess huge sets of MRI records and draw the attention of medical staff upon suspected positive cases. This paper proposes a machine learning solution based on binary decision trees and random forest technique, trained to provide accurate segmentation of brain tumors from multispectral MRI volumes. The current version of our system was trained and tested using all 220 high-grade tumor volumes from the MICCAI BRATS 2016 database. Image records were preprocessed to attenuate the effect of relative intensities in the MRI data, and to extend the feature set with neighborhood information of each voxel. The output of the random forest is also validated for each voxel, according to labels given to neighbor voxels. The achieved accuracy is characterized by an overall mean Dice score of 80.1%, sensitivity 83.1%, and specificity 98.6%. The proposed method is likely to detect all gliomas of 2 cm diameter.

Random forest feature selection approach for image segmentation

In the field of image segmentation, discriminative models have shown promising performance. Generally, every such model begins with the extraction of numerous features from annotated images. Most authors create their discriminative model by using many features without using any selection criteria. A more reliable model can be built by using a framework that selects the important variables, from the point of view of the classification, and eliminates the unimportant once. In this article we present a framework for feature selection and data dimensionality reduction. The methodology is built around the random forest (RF) algorithm and its variable importance evaluation. In order to deal with datasets so large as to be practically unmanageable, we propose an algorithm based on RF that reduces the dimension of the database by eliminating irrelevant features. Furthermore, this framework is applied to optimize our discriminative model for brain tumor segmentation.

Gaussian refinements on Gabor filter based patch descriptor

Recent research in object detection tends to put an accent not only on global object methods, but concentrates mostly on object parts and the relationship between them. One of the most widespread part-based object model was proposed by Felzenszwalb et al. [1]. Such systems can be divided into three main parts: the detection of interest points, the development of adequate local descriptors and the object model. This article deals with the most important phase: the elaboration of local descriptors. We have therefore created a patch descriptor based on two-dimensional Gabor filters. The idea of the descriptor thus developed is to select only a few from the multitude of definable Gabor filters: those most adequate for a given object part. In our previous works, we designed a response-map that played the role of the local descriptor, based on the above-mentioned filters and the GentleBoost learning algorithm [2] or the SVM classification method [3]. In this paper we propose an improvement to the filter selection process which considers not only the magnitude of the complex Gabor filter responses, but the real and imaginary parts and their statistical distribution as well. For this purpose, we have created an RBF Neural Network able to learn the statistical distribution of Gabor filter responses. This network improved the selection procedure of the most suitable filters for a given image patch. The idea of using the RBF NN has been suggested by several authors [4, 5, 6] whose systems are based on the Gaussian distribution of Gabor filters. In conclusion, we have compared the above-mentioned three methods – GentleBoost, SVM and RBFNN – and have deduced that the combination of Gaussians characterizes the patch better than just the magnitude value of the complex responses.

Biometric identification based on feature fusion with PCA and SVM

Biometric identification is gaining ground compared to traditional identification methods. Many biometric measurements may be used for secure human identification. The most reliable among them is the iris pattern because of its uniqueness, stability, unforgeability and inalterability over time. The approach presented in this paper is a fusion of different feature descriptor methods such as HOG, LIOP, LBP, used for extracting iris texture information. The classifiers obtained through the SVM and PCA methods demonstrate the effectiveness of our system applied to one and both irises. The performances measured are highly accurate and foreshadow a fusion system with a rate of identification approaching 100% on the UPOL database.

A Self-tuning Possibilistic c-Means Clustering Algorithm

Most c-means clustering models have serious difficulties when facing clusters of different sizes and severely outlier data. The possibilistic c-means (PCM) algorithm can handle both problems to some extent. However, its recommended initialization using a terminal partition produced by the probabilistic fuzzy c-means does not work when severe outliers are present. This paper proposes a possibilistic c-means clustering model that uses only three parameters independently of the number of clusters, which is able to more robustly handle the above mentioned obstacles. Numerical evaluation involving synthetic and standard test data sets prove the advantages of the proposed clustering model.

Biometric System Based on Registration of Dorsal Hand Vein Configurations

In this paper we present a biometric system based on dorsal hand vein recognition. The preprocessing steps are tuned for image similar or captured with the same scanner as used for the creation of NCUT database. Image quality was improved according to the segmentation method applied. A coarse segmentation technique based on ordinal image encoding has been proposed to determine the significant parts of the vein skeleton. The vein skeleton obtained is the basis of an accurate image registration. The current work shall prove that the geometric attributes of the segmented vascular network are a solid basis for the dorsal hand vein registration process. The designed authentication system is based on the similarity of registered images applying the k-NN classification. A novel and promising similarity method capable of measuring the distance between two point sets, which have comparable visual aspects, has been introduced. The system was evaluated on the NCUT database. The experimental approach shows that the geometric attributes proposed can reach high performances (near 100% accuracy on the considered database).

Detecting the eye and its openness with Gabor filters

Drivers fatigue is the major cause of traffic accidents all over the world. Advanced image processing technology processing the stream obtained from infrared cameras is able to supervise blinking rate and at the same time drowsiness of the vehicle driver. Such a system may warn not only the tired person, but also the passengers, whom the driver takes responsibility for. In this article we present a method of determining the openness of the eye from still images. The method proposed is based on the eye detection presented in our previous work [1]. The eye is detected by using some fine-tuned Gabor filters, specially developed for this purpose. In order to speed up eye detection, the well-known Viola-Jones [2] face detection method is used, implementing self-trained classifier obtaining lower false positive detection rate [3]. The created final classifier is able to detect more pixels of the open eye and compares it to the eyelid marked with very few pixels, meaning a closed eye.

Bias field inhomogeneity measurements

Magnetic resonance imaging (MRI) is affected by intensity inhomogeneity where the illuminated areas alternate with shadow areas. The phenomenon of inhomogeneity is barely noticeable by the human observer, but in the field of automated image segmentation or registration, the unwanted intensity variations can cause significant errors. An important step in image processing is the evaluation of the inhomogeneity and the adequate correction of this artifact. In this paper we have proposed to measure the inhomogeneity with the most well-known quantitative formulas. We shall also put forward two new measurement methods: one for direct measurement and one for indirect measurement. The first method measures the smoothness of the bias field ratio, regardless the used correction algorithm. The second defines a procedure and an image function for inhomogeneity evaluation. The values obtained point the inhomogeneity level out and suppress the white noise better than the usual algorithms. We used simulated MR images obtained from the Brain Web site in the quantitative and comparative evaluations. The results obtained, when compared to existing methods, show a significant confidence level for the proposed measurements.

An Atlas Based Performance Evaluation of Inhomogeneity Correcting Effects

Abstract In automated image processing the intensity inhomogeneity of MR images causes significant errors. In this work we analyze three algorithms with the purpose of intensity inhomogeneity correction. The well-known N3 algorithm is compared to two more recent approaches: a modified level set method, which is able to deal with intensity inhomogeneity and it is, as well, compared to an adaptation of the fuzzy c-means clustering with intensity inhomogeneity compensation techniques. We evaluate the outcomes of these three algorithms with quantitative performance measures. The measurements are done on the bias fields and on the segmented images. We consider normal brain images obtained from the Montreal Simulated Brain Database.