Speech refers to a description of the preparatory work for the session of the risk analysis. These works cover: identification of standards and requirements, selection of an appropriate risk analysis procedures for staff working in different countries, determine the type of documentation for the software, preparation of input documents for risk analysis.

The report presents the background of initiation the research on a problem of objective and repeatable assessments of large number of ultrasound images. Currently used methods, based on estimation of the inflammation area and blood flow in this area, are presented together with their limitations. Potential problems related to influence of human factor on these assessments are shown. MEDUSA project has been proposed as a remedy for these problems. The start-up phase of the project is presented with its challenges related to cooperation of several interdisciplinary teams from multiple locations around the world. The progress of the project and evaluation of the activities accomplished so far (scientific, educational, organizational/logistics and administrative) are presented.

We suggest a fast multi-scale edge-detection scheme for medical ultrasound images. The edge-detector is based on well-known properties of the continuous wavelet transform. To achieve both good localization of edges and detect only significant edges, we study the maxima-lines of the wavelet transform. One can obtain the maxima-lines between two scales by computing the wavelet transform at several intermediate scales. To reduce computational effort and time we suggest a time-scale filtering procedure which uses only few scales to connect modulus-maxima across time-scale plane. The design of this procedure is based on a study of maxima-lines corresponding to edges typical for medical ultrasound signals. This study allows us to construct an algorithm for medical ultrasound signals which meets the demand for speed. Our results show that the proposed algorithm effectively detects major features in such signals, including edges with low contrast.

The role of image registration in a synovitis detector is to identify the parts in ultrasound images such as bone and joint that will be used to guide the search for a possible inflammation region. The registration task is formulated as a problem of finding a correspondence between a set of image features and an articulated model that includes a set of parts and descriptions of geometric relations among the parts. The image features are the regions in an image that result from an application of
a group of detectors such as joint, bone, skin and tendon detectors. While such individual detectors are imperfect, the registration method combining their results with learned relational constraints has a high likelihood of correctly identifying the parts. The registration method is based on an objective function (cost) that includes global and local rigid transformations. It proceeds by searching for a correspondence between the model and the image feature set that minimizes the cost. The models are inferred from a set of training images with annotations, by a combined supervised and unsupervised learning.

A learning approach to detecting a joint location in ultrasound images is being developed, using point feature descriptors. The training and test sets consist of images with the joint regions identified. An image point feature descriptor such as SURF is used as the feature vector for a classifier. A pixel classifier is constructed by training multiple simple classifiers, including k-nearest neighbor, nearest descriptor cluster, and SVM, from which a composite classifier is constructed using an ensemble learning method such as boosting. To increase the computational efficiency, the pixels where a descriptor is computed are initially screened by applying an image point feature detector. The final joint detector is the result of clustering the pixels classified as "joint region".

A detector is being developed for identifying linear forms in ultrasound images such as bone, skin and tendon. The detector is a trainable classifier which applied to a stack or vector of images. The components images of the vector are the results of passing the input image through filters which enhance linear image features such as edges and ridges, including the first and second derivative filters, Laplacian and threshold operators. The classifier is represented by a vector of convolution kernels that correspond to the component images, and associated weight factors. The classification function is a weighted sum of normalized convolutions of the component images and their corresponding component kernels. The kernels and weights of the classifier are computed from a training set of ultrasound images with identified bone, skin and tendon forms, using clustering and supervised learning methods. The ITK toolkit is used for image processing operations.

MEDUSA annotation editor is being developed to support adding annotations such as labels, descriptions and region outlines to ultrasound images, that will form a training set for a learning synovitis detector. The presentation starts from base requirements, presents continuous integration process adapted to the project, existing and planned application features and rules of iterative deployment and requirements management established with a team from Norway. It mainly refers to  first tasks from WP3 – Prototype Software Development in particular: T.3.1. Database system construction and T3.2. Development of annotation editors. For performance reasons and flexibility it has been chosen C++ programming language. Target software will run in multiplatform environment – in particular Linux and Windows platforms with possible extensions to mobile platforms. Target application will have modular structure, and will enable a simple extension of its capabilities with dedicated plugin system. Exchanging/publishing plugins between users will be supported.

The report presents findings of the review of data mining tools which may be applied for medical images evaluation. However, the problem of joint synovitis activity analysis from medical ultrasound and Power Doppler examinations in not supported by any known data mining tools but same of its sub-problems are supported. This includes general purpose data mining tools that allow for image mining activities available in the form of extension libraries. The most relevant to MEDUSA research area extensions allow for solving generic problems in image segmentation, feature extraction, pattern detection, and image classification. This report presents a few free accessible image mining solutions which are available for popular data mining tools.  Three widely accessed data mining environments were considered: Weka, RapidMiner, and KNIME. The particular attention was given to problem how to adopt general purpose data mining tools to support medical images analysis in the research context of MEDUSA project activities.