International Conference on Computer Systems in Medicine and Health - spring 2016
Overview of tensor methods in image processing and computer vision
Image processing and computer vision require processing of large amount of multi-dimensional data. Examples are face recognition, image synthesis, video analytics for surveillance, medical data analysis, to name a few. Processed data depend on many factors and the classical vector based methods frequently become unsatisfactory since they do not exploit full information contained in data and their structure. However, recently developed tensor based methods allow data representation and analysis which directly account for multi-dimensionality. In this talk we will present an overview of tensor data representation, tensor decompositions, as well as pattern recognition with tensor based classifiers with applications in image processing and computer vision. Practical aspects and implementation issues will be also presented.
Medusa project: The objectives and initial results
Konrad Wojciechowski, Bogdan Smolka, Rafał Cupek, Adam Ziębiński, Karolina Nurzynska, Marek Kulbacki, Jakub Sagen, Marcin Fojcik, Paweł Mielnik, Sebastian Hein
SUT, PJATK, HiSF, HF, ITAM
This article describes the research goals undertaken in the MEDUSA project, which concentrates its interests on the problem of automatic estimation of finger joint inflammation activity state using the information present in ultrasonography imaging (USG). The stages which were assumed to proceed were detailed. Some results achieved during the initial part of the realisation were discussed.
A patch-based region growing segmentation of finger joint synovitis in ultrasound images
Ultrasonography has proved its usefulness in the evaluation of joint inflammations caused by rheumatoid arthritis. The illness severity is scored by human examiners based on their experience, but some discrepancies in the final diagnosis and treatment frequently occur. Therefore, the main aim of this work is the elaboration of an automatic method of the localization and assessment of joint inflammation level in ultrasound images.
In this work the seeded region growing method is applied for synovitis region segmentation. The region growing method is a technique that extracts a region from the image using some predefined criteria of similarity between initially selected point and the pixels in its neighborhood. The seed points are placed automatically as the darkest patch within a small region between two detected finger bones. The bones are localized using confidence maps, which are directly related with the likelihood of ultrasound transmission. The region affected by synovitis is found using the adopted criterion of homogeneity based on patch to patch similarity measure. The obtained results show a high accuracy in comparison with the annotations prepared by the expert.
An algorithm for joint and bone localization in usg images of rheumatoid ARTHRITIS
The assessment of degree of joint inflammation using USG imaging is a challenging task. The rating of disease severity and the indicated areas of synovitis can differ even between experts. Therefore, the evaluation of images by a computer program may be a vital solution for objective assessment of disease progression. This is the main goal of MEDUSA project. In this paper an original pipeline, developed for precise estimation of bone line and joint location in USG images, is proposed. The main part of the algorithm utilizes a novel and simple technique of image filtering, which allows for highlighting and, subsequently, determining the area of bones. The search for joint position is based on the analysis of local minima in estimated bone line. The presented routines were fine-tuned be several parameters to minimize the distance between algorithm’s outcomes and the reference markings indicated by an expert. The presented set of methods provides a basis for further classification tools, which will aim at identification and rating of the degree of joint inflammation.
Adaptive human skin segmentation in digital images
Michał Kawulok and Jakub Nalepa
Segmenting human skin regions in colour images is an important and challenging task of computer vision that has drawn research attention over the years. A general goal here is to determine whether human skin appears in a presented image or video sequence, and subsequently to precisely extract the entire skin region. Potential applications of skin detection and segmentation are quite wide and they include gesture recognition and human-computer interaction, objectionable content filtering, feature extraction for content-based image retrieval, skin lesion detection in medical images, and many more. Accuracy of colour-model-based skin detectors is limited due to high variance and low specificity of the skin colour. Possible remedies to this problem include involving spatial features, as well as creating adaptive colour models which are dynamically tuned to a presented case. In this presentation, we will focus on the developed adaptation schemes that make it possible to adapt skin colour models on the fly to mitigate the problems resulting from high variance of the skin colour. In particular, we will discuss our approach to determine the adaptive seeds for geodesic distance transform, which allows for combining the benefits of adaptive skin colour modelling with the spatial analysis of skin-like pixels. These adaptive seeds are extracted either from the facial regions detected based on human face geometry, or retrieved automatically from skin-presence probability maps obtained using a global skin colour model. We will present the experimental results which demonstrate high competitiveness of our approach compared with existing state-of-the-art methods.
CFD analysis of blood flow within aorta and its thoracic branches of patient with coarctation of aorta
Marek Rojczyk, Ziemowit Ostrowski, Wojciech Adamczyk, Bartłomiej Melka, Dominika Bandoła, Maria Gracka, Artur Knopek, Professor Andrzej J. Nowak
The untreated coarctation of aorta CoA (narrowing of the descending aorta opposite to the site of the ductus arteriosus insertion) causes premature death of the patients. The negative effect on the cardiovascular system comes from exposure of the upper part of the human body to hypertension and blood flow disturbances with all consequences. According to the American College of Cardiology and American Heart Association guidelines for treatment patients with congenital heart disease indications for the therapeutic intervention in CoA require invasive blood pressure measurement before and after the narrowed site. A non-invasive assessing of the coarctation severity solely on the data from imaging modalities like CT or MRI supported with blood flow modelling using computed fluid dynamics would be very beneficial for the clinical practice.
In the work pulsatile blood flow in real geometry of aorta and its thoracic branches of 8-year old patient was analysed. The Computational Fluid Dynamics (CFD) analysis has been carried out using commercial code ANSYS Fluent (Ansys Inc., USA). Inlet and outlet boundary conditions were implemented using User Defined Function (UDF) technique . As inlet condition velocity profile was used to mimic flow conditions during human cardiac cycle. The outflow conditions was set to pressure outlet type, where current outflow pressure is calculated using lumped parameter model (LPM). The LPM was based on electrical analogy, where systemic arterial system (hydraulic circuit) is described by 3-element Windkessel model. The blood was modelled as non-Newtonian fluid, flow was treated as single-phase flow.
Comparison of three in silico models of heat transfer in human body
Piotr Buliński, Ziemowit Ostrowski, Wojciech Adamczyk
Heat transfer and temperature distribution is important not only for thermal comfort but it also can be adopted to diagnose of skin lesions, with special interest in malignant melanoma identification. The temperature distribution and heat flux on skin can be examined using thermographic camera. Brass compress is placed on the forearm skin to perform mild skin cooling process. After short period of time compress is removed and recovery temperature distribution can be recorded.
Investigated numerical model includes the heat transfer equation in the living tissues, in form of well-known mathematical formulation proposed by Pennes’. In basic form, this bioheat equation covers heat conduction and two source terms: perfusion and metabolic. While in more sophisticated approaches, like one suggested by Fiala, additional models of both passive and active thermoregulation are applied. Those models are validated against measurement data within current study. Used properties of human tissues hard to measure in vivo, are based on well-established literature sources. To test their influence on model, the sensitivity analysis has been performed. The comparison of bioheat transfer models is presented and discussed.
Extracting Biomarkers from Dynamic Contrast Enhanced Images
Jakub Nalepa and Michał Kawulok
Imaging technologies have developed rapidly over the past decade proving to be valuable and effective tools for diagnosis, evaluation and treatment of many conditions, especially cancer. Dynamic contrast enhanced (DCE) imaging using computed tomography (CT) or magnetic resonance (MR) has been shown particularly effective and has been intensively studied to allow for assessing the vascular support of various tumours and other tissues. DCE biomarkers were proven to be correlated with physiological and molecular processes which can be observed in tumour angiogenesis (morphologically characterised by an increased number of micro-vessels, which are extremely difficult to image directly). Therefore, DCE biomarkers can robustly assess tumour characteristics and stage, and provide an independent indicator of prognosis, enabling risk stratification for patients with cancer. In this presentation, we will discuss the challenges behind extracting biomarkers from dynamic medical images. The process of the DCE analysis involves acquiring time series images and investigating temporal changes of injected contrast material (tracer) attenuation in vessels and tissues. The process of quantifying such changes encompasses several steps, including 4D image segmentation, image registration, and DCE analysis – every step in the processing pipeline is pivotal and affects the final numerical scores (extracted biomarkers). We will discuss the possibilities of applying various machine learning techniques and tools to boost the performance of the DCE analysis. Finally, we will go through the most important practical issues that need to be addressed in our computational system for extracting and post-processing of DCE biomarkers that is being developed at the moment. Satisfying these requirements is quite challenging, but it is crucial for bringing such systems into a day-to-day clinical practice (thus into the FDA/21 CFR 11 regulated environment).
Modelling cerebrovascular network and blood microcirculation
Krzysztof Psiuk-Maksymowicz, Marcin Mazur, Damian Borys, Jarosław Śmieja
The aim of this work was creation of the vascular network of cerebrum and simulation of blood flow through it.
The automatic generation of brain vascular network contains arterioles, capillaries and venules. Such vessels differ from each other in diameters and lengths. The positions of the vascular network nodes are calculated based on quaternions. Developed algorithm generating vascular structures exploits physiological parameters of a brain. Implementation allows to simulate pressures and blood flow for each segment of the network.
We analysed the model in terms of physiological correctness of generated vascular network structure. Application of particular model parameters resulted in creation of incorrect network structure, for example structure which does not cover whole computational space.
Performed simulations properly map the dependences between vessel diameters and blood velocities, and vessel diameters and blood pressure.
Presented model may have use wherever characteristics of microcirculation are important. The potential application of the model are studies of atherosclerotic lesions in silico as well as studies of tissue perfusion which do depend on vascularisation.
Overview of available open source PACS framework
Stanisław Wideł, Andrzej Wideł, Dominik Spińczyk
SUT, Hospital in Rybnik
The development of medical services requires designing of informatics tools that are necessary to generate progress in the existing scientific disciplines in the field of healthcare. The article presents an overview of the available informatics systems for archiving medical images in open-source solutions, taking into account popular criteria for software evaluation.
CASE STUDY OF orthopedic PRE-SURGICAL PLANNING USING 3D PRINTED MODEL
Stanisław WIDEŁ, Agnieszka SZCZĘSNA, Andrzej WIDEŁ
SUT, Hospital in Rybnik
Traditionally, orthopedic surgeons operations are planned based on the CT and MRI images of patients. While these images can illustrate a patient’s organ from different angles, they might not show all injuries that could cause possible complications. Also only visualizations with 3D models are not sufficient because do not allow to fit and adjust necessary tools and components.
The case study of the pre-surgical planning with use of printed 3D model is presented. On the basis of this case has been described in detail all the stages of preparation and printing 3D model.
The 43-year-old patient was admitted to the Department of proximal tibia fracture (type VI by Schazker classification). Injury was as a result of a fall of horse. With the admission diagnostic imaging was performed including CT scan of the knee and proximal tibia. With use of PACS sewer the DICOM files were stored and available in other units. The PACS sewers are oriented to the organized structure of patient data including examination with some sequences.
With use of the Silcer software with manually adjusted parameters of segmentation the 3D model of bone was prepared. Model is the triangle surface mesh. A 3D model can be outputted on a 3D printer if the model is a 2-manifold triangular surface mesh, that means there are no holes and non-manifold edges or faces. Preprocessing of 3D model requires error correction, decimation of mesh, holes closing, scaling and positioning on the printer ground plane. To do this step we use MeshLab software. Properly prepared mesh saved in STL file can be send layer by layer in G-Code to the printing device. Adjusting model to specific 3D printer, selecting and inserting addition supports and adjusting the printing parameters to material is available in Slic3r software. This software also generates G-Codes for control the printing process. We use small and cheap PrintBot with PLA filament.
With use of printed 3D model was planned the degree of correction required to achieve congruence articular surface, reproduce limb length and axis of rotation and augmentation of bone defects by transplantations.
Study and adjustment of dermoscopic image preprocessing algorithms for lesion border detection
Damian Borys, Magdalena Szeremet, Krzysztof Psiuk-Maksymowicz, Ziemowit Ostrowski, Mariusz Frackiewicz
Dermoscopy is a basic technique in skin melanoma diagnostics. The goal of our work was an initial preprocessing of dermoscopic images towards accurate lesion border detection. Dermoscopic images are frequently occluded by air bubbles when immerse liquid is used and patient hairs. Those hairs partially shade the main region of interest that’s why it needs special treatment.
Materials and methods: All algorithms were tested on PH2 images database that contains 200 dermoscopic images, each with a mask of the lesion. Those images were a gold standard for each algorithm. We have developed and tested four algorithms of image preprocessing and lesion contouring. Each of those algorithms requires some parameters so why we have optimised each of proposed algorithm using lesion mask from PH2 database and Jaccard index as a measure of similarity of both sets. Four algorithms were proposed: MS - algorithm using mean shift clustering, HE - algorithm using histogram equalisation, TTH - algorithm using the top-hat transform, PCA - algorithm using principal component analysis. Each algorithm has a multi-step structure, beginning from conversion to grayscale. In all of them, one of the crucial steps is morphological algorithms. Parameters in each algorithm were determined using Jaccard index and images with lesion masks from PH2 database. Optimal parameter sets for each algorithm were used to compare those algorithms. Each image from the database was analysed and Jaccard index was calculated. Simple statistical analysis of indexes was used to compare proposed algorithms in term of their accuracy.
Simon Buvarp, Ørjan Malkenes
The presentation is about our senior bachelor project for Helse Førde. The purpose of this project is to reduce the time spent looking for the destinations of different departments at Førde hospital, since this is a big and complex building. We will explain how we are going to solve the problem, using an android application which helps you to find the way. The application will use QR-scanning, search functions and Dijkstra algorithms. http://www.pathfinder-hisf.no/
Rehabilitation with robotic arm
Marius Blom, Børge Godejord, Hans Olav
We are looking at the possibility of using a robotic arm in rehabilitation of stroke patients. Stroke is an increasing problem in Norway and we see a potential for robotic arms to operate in this field in the future. We have tried to replicate some exercises commonly used in the rehabilitation of patients, and looked at some advantages a robot might bring. http://hisf.heime.nu/
Some challenges in implementation of ICT in medicine
Marcin Fojcik, Joanna Gałek
Engineers, physicians and nurses perceive reality in different way. This presentation will show the most common problems observed during implementation of MEDUSA and others projects carried by Sogn og Fjordane University College and Helse Førde.
Practical approach to ultrasound synovitis classification
Semiquantitative ultrasound with power Doppler is a reliable and widely used method of assessing synovitis. Synovitis is estimated by ultrasound examiner using the scoring system graded from 0 to 3. Activity score is estimated on the basis of the examiner’s experience or standardized ultrasound atlases. The method needs trained medical personnel and the result can be affected by a human error. We present method used to decrease discrepancies between observers. We present preliminary results experts classification of control data for the MEDUSA project.
Automatic search of optimal process configuration with applications to anatomic feature detectors and ultrasound image segmentation
Detection of desired target regions in an image is a multistage process, where an optimal configuration of a substantial number of parameters must be determined for each stage. For example, a detector of bones and skin in an ultrasound image is divided into the stages of representative image features extraction and classifier learning. Another example, an image segmentation method is divided into a rough contour detection stage and a contour propagation stage. For practical reasons, different stages may be implemented as separate processes in different languages, require different tools, and even different operating systems, which makes the parameters search more difficult. As the exhaustive search over all possible parameters is not possible in the practice, an often used technique is an intuitively guided manual search of good parameter configurations, which is time consuming and inefficient. This presentation describes a method of searching a large space of parameters based on a customizable protocol for definition of complex experiments, automatic iterative execution of a large number of test cases, and filtering and visualization of results. The presented method enables the use of different languages, tool sets and operating systems for image processing and machine learning, and offers a single consistent optimization process for all stages of detection.
Cloud based evaluation system for synovitis estimation methods and modules
The paper presents a prototype of a cloud based evaluation system designed for synovitis estimation. This web based tool allows collecting, processing and visualization of detected anatomical structures and synovitis area. This system is both, a prototype of a software system intended to be used by medical personnel to help in diagnosis of rheumatic diseases, and an environment for evaluating the synovitis estimator by medical experts. It can be used for automatic finding and tagging of elements such as joints, bones or skin and tissue. Users are able to upload USG data and receive the results in the form of tagged pictures. The presented cloud based solution improves the effect of data evaluation.
Post-processing results of bone and skin detectors for synovitis estimator
A bone and skin detector is a preparatory image processing stage in a process of estimation of the degree of synovitis from an USG image. This detector produces points considered to be a part of either a bone or skin. Further processing stages require that connected regions formed from such output points are represented as polygonal chains. Some of the detected regions are noticeably in error, such where a small patch labeled “bone” lies next to a large patch labeled “skin” and vice versa. Correcting such errors and forming polygonal chains is done in a post-processing phase, which consists of three stages: cleaning output data, clusterization and linearization. The cleaning stage, based on sibling voting, corrects a large number of false detections. The clusterization turns a set of detected points into a connected region. The linearization stage uses iterative line fitting and subdivision to fit polygonal chains to connected regions, which is the data format needed by the next stage of the synovitis estimator.
Evaluation of two methods of segmentation of USG images for synovitis region detection
Determination of a synovitis degree from USG image of a human finger joint depends on extraction of a synovitis regions from the image. In this work we present evaluation of two methods approaching this problem, the first based on a classical image segmentation, the second based on a trained classifier of image pixels. The classical segmentation approach applies a spatial-color based clusterization of images and a mixture of watershed and flood fill methods. The pixel classification method treats the synovitis region extraction as a supervised learning problem. In the training phase, each image pixel was assigned one of two class labels: “in synovitis” or “outside of synovitis”. Several local neighborhood descriptors like SURF, BRISK, FAST, ORB were used as feature vectors for pixels, and Support Vector Machine was used for learning the classifiers.
For evaluation of the synovitis extraction methods we created a measure based on comparison of the area and boundary of the detected and annotated synovitis regions, averaged over all images in the test set. The presented results are better for the learning based method than for the classical segmentation, but they are still below our expectations.
Unsupervised learning of finger joint models from USG images
Bones, skin or the joint center can be identified in an ultrasound image using a structural model, consisting of parts representing the bones and skin as polygonal chains and the join as a point. The parts are localized in an image by finding a correspondence between a structural description computed from an ultrasound image with the structural model. The structural models used for bones, skin and joint recognition are collected in a library, where each model represents a class of joint images. All the library models together are intended to cover the range of structural descriptions that occur in training images.
A method of unsupervised learning is described, which is used to build a library of structural models, by clustering structural descriptions extracted from the annotations of images in the training set. This method uses a distance-like measure, based on an objective function that quantifies the discrepancy between structural descriptions. A correspondence between a structural model and a structural description is obtained using a search that finds the correspondence which minimizes the objective function.
Resolving conflicts between human skeleton representation in virtual reality and user's pose
The power of virtual reality is only as great as the VR is immersive. Applications ranging from entertainment to battlefield simulations to telemedicine, all rely on the perception that the user is present at virtual location and therefore is able to interact with it using any kind of interface. To obtain best results, the interface has to be as natural as possible, and the user's own body movement is most natural way of controlling his virtual avatar. Data from sensor-based suits can be used to obtain human motion, however lack of real counterparts to the virtual objects that user interacts with can lead to the differences between user's pose and it's virtual representation which is restricted by collisions with virtual environment.
The issue at hand is thus how to manipulate avatar's pose in order to remove any differences between it and user's pose in a way that will not break the immersion. In presented approach, range of motion data acquired from marker-based motion capture system is used to obtain anthropometric constraints for human body as well as a frame by frame quaternion-based rotation paths between different poses for each segment and joint. Minimization of energy based on angular velocity in joints is used in order to estimate the most natural way for virtual representation to seamlessly return to the pose enforced by the user while still reacting to his movement during the state of discrepancy, with no instantaneous, unnatural changes in skeleton's segments orientation.
Presented approach reduces discrepancies between user's pose and it's virtual representation, which allows for immersive use of virtual reality headsets with motion capture suits based on inertial measurements units.