Search

In the present study we evaluate the performance of various training schemes for the locally recurrent probabilistic neural network and seek for advantageous tradeoffs between required training time and classification accuracy. Specifically, we consider training schemes which make use of a simple incremental procedure for adjusting sigma, as well as methods based on particle swarm optimization or differential evolution in different configurations. The experimental evaluation was carried out in common experimental protocol based on the Parkinson speech dataset. The experimental results show that with a proper training configuration a high accuracy can be achieved even with limited training data.

This paper presents a method for creation of computational models of breast lesions with irregular shapes from patient Digital Breast Tomosynthesis (DBT) images or breast cadavers and whole-body Computed Tomography (CT) images. The approach includes six basic steps: (a) normalization of the intensity of the tomographic images; (b) image noise reduction; (c) binarization of the lesion area, (d) application of morphological operations to further decrease the level of artefacts; (e) application of a region growing technique to segment the lesion; and (f) creation of a final 3D lesion model. The algorithm is semi-automatic as the initial selection of the region of the lesion and the seeds for the region growing are done interactively. A software tool, performing all of the required steps, was developed in MATLAB. The method was tested and evaluated by analysing anonymized sets of DBT patient images diagnosed with lesions. Experienced radiologists evaluated the segmentation of the tumours in the slices and the obtained 3D lesion shapes. They concluded for a quite satisfactory delineation of the lesions. In addition, for three DBT cases, a delineation of the tumours was performed independently by the radiologists. In all cases the abnormality volumes segmented by the proposed algorithm were smaller than those outlined by the experts. The calculated Dice similarity coefficients for algorithm-radiologist and radiologist-radiologist showed similar values. Another selected tumour case was introduced into a computational breast model to recursively assess the algorithm. The relative volume difference between the ground-truth tumour volume and the one obtained by applying the algorithm on the synthetic volume from the virtual DBT study is 5% which demonstrates the satisfactory performance of the proposed segmentation algorithm. The software tool we developed was used to create models of different breast abnormalities, which were then stored in a database for use by researchers working in this field. ispartof: PHYSICA MEDICA-EUROPEAN JOURNAL OF MEDICAL PHYSICS vol:57 pages:80-87 ispartof: location:Italy status: published

The current contribution presents an approach for creation of complex computational breast phantoms, with included breast abnormalities. In-house developed and validated software tool, called BreastSimulator, is used to create different models of healthy breasts, by varying parameters like shape, size, duct tree features, Cooper ligaments, skin, etc. Separately, different 3D tumour models are constructed from segmented tumour regions on sets of real tomosynthesis images from real patients. A semi-automatic algorithm was developed to segment the abnormalities in the images. The algorithm applies a series of image processing operations and region-growing techniques with minimal interaction from the user to find the areas of the lesions. The flat tumour images are then correctly stacked and stored in a 3D voxel matrix, matching the desired resolution. Relevant information about the lesion (voxel size, matrix size, geometrical centre, etc.) is also saved. Selected tumour models are then inserted at various locations into the computational models of the healthy breasts. Examples of simulated mammographic and tomosynthesis images from such breast phantoms are presented to illustrate the potential of the proposed approach. Combining mathematical breast modelling with segmented real 3D tumour shapes results in the creation of realistic breast phantoms. The approach allows implementation of multiple scenarios and unlimited number of cases, which can be used for further software modelling and investigation of breast imaging techniques.

The Locally Recurrent Probabilistic Neural Network (LRPNN) consists of an input layer, three hidden layers and an output layer. The first two hidden layers are derived from the original PNN, while the third layer referred as recurrent layer is capable to model correlations within temporal sequences of observations. In the present study, we investigate the feasibility of FPGA-based implementation of the locally recurrent layer of LRPNN. An important consideration due to the specifics of this architecture is the use of modules with very high precision in the hardware design. Although expensive in terms of available resources in the FPGA chip, this is necessary, in order to compensate for the added error of quantization due to the multiple feedbacks from neurons in the neural network. The weights for the recurrent layer of the LRPNN are automatically computed from the available training data and translated into the hardware design. The experimental evaluation was carried out on the DEAP database, where two classes of emotional states were considered. The design makes use of a computed short-term energy from a 32-channel electroencephalographic (EEG) signal as an input. Results of an extensive experimental validation show that there is approximately one percent difference between the accuracy achieved with CPU-based software and FPGA-based hardware implementation of the LRPNN.

A common approach in the development and improvement of diagnostic imaging techniques is the use of anthropomorphic phantoms. These phantoms can be physical or computational. In this study the creation of computational breast phantoms with included pathological formations is presented. The creation of the realistic phantoms is achieved by utilizing real patient data in the form of tomographic images. The 3D tumour models are generated by segmenting the regions containing tumour formations in the patient images. The segmentation is performed with a developed software tool based on a semi-automatic algorithm, which makes use of a series of image processing and region growing techniques. The software tool also provides the user an opportunity for corrections after the automated segmentation. Then the acquired flat images are stacked in a 3D voxel matrix. Creation of the computational healthy breast model as well as the compression procedure is achieved with a software tool called BreastSimulator. The healthy breast model and the segmented tumour formation are then interactively combined with a software tool called XRAYImagingSimulator. While the user can select a location for the tumour formation, also an automatic software processing is applied for integration between the two computational models. The simulation procedure for acquiring tomographic images from the created realistic breast phantom with included tumour formation is performed with the XRAYImagingSimulator software tool. Finally, the acquired simulation images are reconstructed with a software tool called FDKR. The combination of mathematical models of the breast and tumour models segmented from real patient data leads to the creation of realistic breast phantoms, which can be used in X-ray imaging simulation studies. The presented approach gives an opportunity for generation of multiple cases of breast cancer; thus allowing for further progress in already existing software models and techniques in diagnostic imaging. Acknowledgements This research is supported by the Bulgarian National Science Fund under grant agreement DN17/2. This project also has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 692097.

Hardware implementation of machine learning methods is often considered challenging, however, it brings major benefits, such as speed of operation and energy efficiency. Having in mind these benefits, we evaluate the performance and complexity of four probability density estimators (PDE) that we consider interesting for hardware implementation of the probabilistic neural network (PNN). We report results of a comparative evaluation of these PDEs for three different sizes of the PNN, evaluated in a common experimental protocol that makes use of the SPECT dataset. We conclude that two of the estimators are more appropriate for hardware implementation in FPGA.