Jan Masek

Novel Method for Localization of Common Carotid Artery Transverse Section in Ultrasound Images Using Modified Viola-Jones Detector

This article describes a novel method for highly accurate and effective localization of the transverse section of the carotis comunis artery in ultrasound images. The method has a high success rate, approximately 97%. Unlike analytical methods based on geometric descriptions of the object sought, the method proposed here can cover a large area of shape variation of the artery under study, which normally occurs during examinations as a result of the pressure on the examined tissue, tilt of the probe, setup of the sonographic device, and other factors. This method shows great promise in automating the process of determining circulatory system parameters in the non-invasive clinical diagnostics of cardiovascular diseases. The method employs a Viola-Jones detector that has been specially adapted for efficient detection of transverse sections of the carotid artery. This algorithm is trained on a set of labeled images using the AdaBoost algorithm, Haar-like features and the Matthews coefficient. The training algorithm of the artery detector was modified using evolutionary algorithms. The method for training a cascade of classifiers achieves on a small number of positive and negative training data samples (about 500 images) a high success rate in a computational time that allows implementation of the detector in real time. Testing was performed on images of different patients for whom different ultrasonic instruments were used under different conditions (settings) so that the algorithm developed is applicable in general radiologic practice.

Evolutionary improved object detector for ultrasound images

Object detection in ultrasound images is difficult problem mainly because of relatively low signal-to-noise ratio. This paper deals with object detection in the noisy ultrasound images using modified version of Viola-Jones object detector. The method describes detection of carotid artery longitudinal section in ultrasound B-mode images. The detector is primarily trained by AdaBoost algorithm and uses a cascade of Haar-like features as a classifier. The main contribution of this paper is a method for detection of carotid artery longitudinal section. This method creates cascade of classifiers automatically using genetic algorithms. We also created post-processing method that marks position of artery in the image. The proposed method was released as open-source software. Resulting detector achieved accuracy 96.29%. When compared to SVM classification enlarged with RANSAC (RANdom SAmple Consensus) method that was used for detection of carotid artery longitudinal section, works our method real-time.

Automated localization of Temporomandibular Joint Disc in MRI images

This paper deals with localization of Temporomandibular Joint Disc (TJD) in Magnetic Resonance Images (MRI). Since the contrast of the TJD is quite low when compared to noise ratio when displayed using MRI, its detection is quite complicated. Therefore the method described in this paper are not not focused the disk itself but detect the most significant objects around TJD, which has usually much higher contrast. For the automatic TJD localization asessment, a training set containing 160 training samples (80 positive and 80 negative) were created and published and several approaches were examined to find the best method. The best results were achieved using support vector machine with Gaussian kernel, which achieved 98.16±2.81% accuracy of detection. The creation of the training models for feature extraction and model evaluation was implemented with RapidMiner tool and the IMMI extension. The models created are published at the IMMI extension homepage and they can also serve as a guide to use of the IMMI extension.

An efficient grading algorithm for non-proliferative diabetic retinopathy using region based detection

The paper proposes an image processing algorithm to grade the severity of Non Proliferative Diabetic Retinopathy. For this disease the most important parameter to classify the stage of the disease is the proximity of abnormalities from the centre of Macula. The proposed algorithm provides an efficient grading technique by segmenting the fundus image into specific regions of interest and avoids redundancy in computation. Instead of detecting abnormalities for the whole fundus image, the proposed method emphasizes on the segmented regions for the abnormalities, thereby reducing the computation time significantly. Furthermore, this approach provides a simple and direct method to measure the severity of the disease. This region based segmentation also has the advantage of a mesh lesser computational load making this process suitable for real time applications. The accuracy of this region based segmentation method is more than 80% when tested in a database.

3D brain tissue selection and segmentation from MRI

Magnetic resonance imaging (MRI) is a visualizing method used in radiology that enables viewing internal structures of the body. Using several mathematical methods with data retrieved from MRI it is possible to quantify the brain compartment volume, which has many applications in cognitive, clinical and comparative neurosciences. This paper introduces a new fully automatic method, which can measure the volume of brain tissue using scans obtained from MRI devices. The method introduced in this paper was trained on data taken from 12 patients and the trained result was validated on other independent data obtained from 10 patients and compared to a human experts accuracy. The result achieves 99.407 % +/- 0.062 voxel error accuracy, which is comparable to results achieved by humans (99.540 % + 0.0775) but in a significantly shorter time and without the need of human involvement.

Speeding up Viola-Jones algorithm using multi-Core GPU implementation

Graphic Processing Units (GPUs) offer cheap and high-performance computation capabilities by offloading compute-intensive portions of the application to the GPU, while the remainder of the code still runs on a CPU. This paper introduces an multi-GPU CUDA implementation of training of object detection using Viola-Jones algorithm that has accelerated of two the most time consuming operations in training process by using two dual-core NVIDIA GeForce GTX 690. When compared to single thread implementation on Intel Core i7 3770 with 3.7 GHz frequency, the first accelerated part of training process was speeded up 151 times and the second accelerated part was speeded up 124 times using two dual-core GPUs. This paper examines overall computational time of the Viola-Jones training process with the use of: one core CPU, one GPU, two GPUs, 3 GPUs and 4GPUs. Trained detector was applied on testing set containing real world images.

Supervised video scene segmentation using similarity measures

Video scene segmentation is a process for dividing video into semantically meaningful blocks. This can help e.g. search engines to divide video into better manageable parts and enable more relevant search in video. Unfortunately, scene segmentation is based on the semantic and therefore it is a difficult task for computers. This work is preliminary study involved into supervised video scene segmentation, which is driven by the way how human segments scenes in a movie. Since these video segments represent semantic parts in video, it can be used for better video annotation and also for searching in videos. As a training set, only high quality movies were used and from these movies 100 training samples have been extracted and used for evaluation. Resulting model is a method based on general color layout, Tamura similarity measure and k-nearest neighbors achieving 97.00% accuracy.

Multivariate classification of echellograms: a new perspective in Laser-Induced Breakdown Spectroscopy analysis

In this work, we proposed a new data acquisition approach that significantly improves the repetition rates of Laser-Induced Breakdown Spectroscopy (LIBS) experiments, where high-end echelle spectrometers and intensified detectors are commonly used. The moderate repetition rates of recent LIBS systems are caused by the utilization of intensified detectors and their slow full frame (i.e. echellogram) readout speeds with consequent necessity for echellogram-to-1D spectrum conversion (intensity vs. wavelength). Therefore, we investigated a new methodology where only the most effective pixels of the echellogram were selected and directly used in the LIBS experiments. Such data processing resulted in significant variable down-selection (more than four orders of magnitude). Samples of 50 sedimentary ores samples (distributed in 13 ore types) were analyzed by LIBS system and then classified by linear and non-linear Multivariate Data Analysis algorithms. The utilization of selected pixels from an echellogram yielded increased classification accuracy compared to the utilization of common 1D spectra.

An SVD based zero watermarking scheme for authentication of medical images for tele-medicine applications

The proposed work presents a zero watermarking method for to solve the issue of medical image security for telemedicine, tele-radiology & tele-opthalmology applications. This method provides medical image security for tele-medicine application without tempering the medical image and no loss of clinical information. Local features in the Singular value decomposition (SVD) domain are used to generate a digital binary code (Master Share) for each fundus image. This master share is strategically combined with encrypted patient ID resulting into a secret share. At the diagnosis centre the patient ID can be accurately recovered by the authorized person only on access of the generated Secret Share. The proposed zero watermarking method is tested on the publically available DRIVE dataset of fundus image and results achieved are encouraging in the direction of medical image identification and verification. The proposed work can be used in telemedicine applications where perfect and loss-less identification is required for medical images as this has direct relevance to human life.

Color image (dis)similarity assessment and grouping based on dominant colors

The computer vision connected to image understanding becomes more and more important in everyday life. This paper concerns the image (dis)similarity assessment and grouping. The main contribution of this paper is the method for image (dis)similarity assessment based on dominant colors. The experimental results showed better results than the Direct Pixel Similarity and Color Histograms and method proved to be capable of detecting images similar to the target image.