Marina Velikova

Improved mammographic CAD performance using multi-view information: a Bayesian network framework

Mammographic reading by radiologists requires the comparison of at least two breast projections (views) for the detection and the diagnosis of breast abnormalities. Despite their reported potential to support radiologists, most mammographic computer-aided detection (CAD) systems have a major limitation: as opposed to the radiologists practice, computerized systems analyze each view independently. To tackle this problem, in this paper, we propose a Bayesian network framework for multi-view mammographic analysis, with main focus on breast cancer detection at a patient level. We use causal-independence models and context modeling over the whole breast represented as links between the regions detected by a single-view CAD system in the two breast projections. The proposed approach is implemented and tested with screening mammograms for 1063 cases of whom 385 had breast cancer. The single-view CAD system is used as a benchmark method for comparison. The results show that our multi-view modeling leads to significantly better performance in discriminating between normal and cancerous patients. We also demonstrate the potential of our multi-view system for selecting the most suspicious cases.

Monotone and Partially Monotone Neural Networks

In many classification and prediction problems it is known that the response variable depends on certain explanatory variables. Monotone neural networks can be used as powerful tools to build monotone models with better accuracy and lower variance compared to ordinary nonmonotone models. Monotonicity is usually obtained by putting constraints on the parameters of the network. In this paper, we will clarify some of the theoretical results on monotone neural networks with positive weights, issues that are sometimes misunderstood in the neural network literature. Furthermore, we will generalize some of the results obtained by Sill for the so-called min-max networks to the case of partially monotone problems. The method is illustrated in practical case studies.

Exploiting causal functional relationships in Bayesian network modelling for personalised healthcare

Abstract Bridging the gap between the theory of Bayesian networks and solving an actual problem is still a big challenge and this is in particular true for medical problems, where such a gap is clearly evident. We argue that Bayesian networks offer appropriate technology for the successful modelling of medical problems, including the personalisation of healthcare. Personalisation is an important aspect of remote disease management systems. It involves the forecasting of progression of a disease based on the interpretation of patient data by a disease model. A natural foundation for disease models is physiological knowledge, as such knowledge facilitates building clinically understandable models. This paper proposes ways to represent such knowledge as part of engineering principles employed in building clinically practical probabilistic models. The methodology has been used to construct a temporal Bayesian network model for preeclampsia – a pregnancy-related disorder. The model is the first of its kind and an integral part of a mobile home-monitoring system intended for use in daily pregnancy care. We conducted an evaluation study with actual patient data to obtain insight into the model’s performance and suitability. The results obtained are encouraging and show the potential of exploiting physiological knowledge for personalised decision-support systems.

A probabilistic framework for image information fusion with an application to mammographic analysis

The recent increased interest in information fusion methods for solving complex problem, such as in image analysis, is motivated by the wish to better exploit the multitude of information, available from different sources, to enhance decision-making. In this paper, we propose a novel method, that advances the state of the art of fusing image information from different views, based on a special class of probabilistic graphical models, called causal independence models. The strength of this method is its ability to systematically and naturally capture uncertain domain knowledge, while performing information fusion in a computationally efficient way. We examine the value of the method for mammographic analysis and demonstrate its advantages in terms of explicit knowledge representation and accuracy (increase of at least 6.3% and 5.2% of true positive detection rates at 5% and 10% false positive rates) in comparison with previous single-view and multi-view systems, and benchmark fusion methods such as naïve Bayes and logistic regression.

On the interplay of machine learning and background knowledge in image interpretation by Bayesian networks

OBJECTIVES To obtain a balanced view on the role and place of expert knowledge and learning methods in building Bayesian networks for medical image interpretation. METHODS AND MATERIALS The interpretation of mammograms was selected as the example medical image interpretation problem. Medical image interpretation has its own common standards and procedures. The impact of these on two complementary methods for Bayesian network construction was explored. Firstly, methods for the discretisation of continuous features were investigated, yielding multinomial distributions that were compared to the original Gaussian probabilistic parameters of the network. Secondly, the structure of a manually constructed Bayesian network was tested by structure learning from image data. The image data used for the research came from screening mammographic examinations of 795 patients, of whom 344 were cancerous. RESULTS The experimental results show that there is an interesting interplay of machine learning results and background knowledge in medical image interpretation. Networks with discretised data lead to better classification performance (increase in the detected cancers of up to 11.7%), easier interpretation, and a better fit to the data in comparison to the expert-based Bayesian network with Gaussian probabilistic parameters. Gaussian probability distributions are often used in medical image interpretation because of the continuous nature of many of the image features. The structures learnt supported many of the expert-originated relationships but also revealed some novel relationships between the mammographic features. Using discretised features and performing structure learning on the mammographic data has further improved the cancer detection performance of up to 17% compared to the manually constructed Bayesian network model. CONCLUSION Finding the right balance between expert knowledge and data-derived knowledge, both at the level of network structure and parameters, is key to using Bayesian networks for medical image interpretation. A balanced approach to building Bayesian networks for image interpretation yields more accurate and understandable Bayesian network models.

Decision trees for monotone price models

Abstract.In economic decision problems such as credit loan approval or risk analysis, models are required to be monotone with respect to the decision variables involved. Also in hedonic price models it is natural to impose monotonicity constraints on the price rule or function. If a model is obtained by a “unbiased” search through the data, it mostly does not have this property even if the underlying database is monotone. In this paper, we present methods to enforce monotonicity of decision trees for price prediction. Measures for the degree of monotonicity of data are defined and an algorithm is constructed to make non-monotone data sets monotone. It is shown that monotone data truncated with noise can be restored almost to the original data by applying this algorithm. Furthermore, we demonstrate in a case study on house prices that monotone decision trees derived from cleaned data have significantly smaller prediction errors than trees generated using raw data.

Learning Bayesian networks for clinical time series analysis

INTRODUCTION Autonomous chronic disease management requires models that are able to interpret time series data from patients. However, construction of such models by means of machine learning requires the availability of costly health-care data, often resulting in small samples. We analysed data from chronic obstructive pulmonary disease (COPD) patients with the goal of constructing a model to predict the occurrence of exacerbation events, i.e., episodes of decreased pulmonary health status. METHODS Data from 10 COPD patients, gathered with our home monitoring system, were used for temporal Bayesian network learning, combined with bootstrapping methods for data analysis of small data samples. For comparison a temporal variant of augmented naive Bayes models and a temporal nodes Bayesian network (TNBN) were constructed. The performances of the methods were first tested with synthetic data. Subsequently, different COPD models were compared to each other using an external validation data set. RESULTS The model learning methods are capable of finding good predictive models for our COPD data. Model averaging over models based on bootstrap replications is able to find a good balance between true and false positive rates on predicting COPD exacerbation events. Temporal naive Bayes offers an alternative that trades some performance for a reduction in computation time and easier interpretation.

2010 Special Issue: Comparison of universal approximators incorporating partial monotonicity by structure

Neural networks applied in control loops and safety-critical domains have to meet more requirements than just the overall best function approximation. On the one hand, a small approximation error is required; on the other hand, the smoothness and the monotonicity of selected input-output relations have to be guaranteed. Otherwise, the stability of most of the control laws is lost. In this article we compare two neural network-based approaches incorporating partial monotonicity by structure, namely the Monotonic Multi-Layer Perceptron (MONMLP) network and the Monotonic MIN-MAX (MONMM) network. We show the universal approximation capabilities of both types of network for partially monotone functions. On a number of datasets, we investigate the advantages and disadvantages of these approaches related to approximation performance, training of the model and convergence.

An integrated reconfigurable system for maritime situational awareness

Nowadays the maritime operational picture is characterised by a growing number of entities whose interactions and activities are constantly changing. To provide timely support in this dynamic environment, automated systems need to be equipped with tools— lacking in existing systems—for real-time prioritisation of the application tasks (missions), selection and alignment of relevant information, and efficient reasoning at a situation level. In this paper, we present METIS—an industrial prototype system for supporting real-time, actionable maritime situational awareness. In particular, we focus on the innovation of METIS, which lies in the employment and integration of several state-of-the-art AI technologies to build the overall systems intelligence. These include reconfiguration of multi-context systems, natural language processing of heterogeneous (un)structured data and probabilistic reasoning of uncertain information. The capabilities of the system have been demonstrated in a proof of concept, which is deployed as a situational awareness plugin in the Tacticos command-and-control platform of our industrial partner. The principles exploited by METIS are giving valuable insights into what is considered to become the next generation of situational awareness systems.

Fully-automated interpretation of biochemical tests for decision support by smartphones

In this paper we present innovative research for the automatic interpretation of biochemical test strip color by a smartphone using image processing techniques. Urinalysis is the current application for these techniques. Our mobile application captures images of the color pads on strips using the camera phone, then analyzes automatically the images within the device itself and compares these against reference color pads to obtain a final classification. As a test scenario we focus on the detection of proteinuria, i.e., leakage of protein into the urine, for the detection of which strips for protein and creatinine are used. We performed an initial laboratory evaluation using specially prepared concentrations to check the accuracy and precision of detection. The results obtained are encouraging and show that the proposed technique has a good potential for the development of cheap, mobile and smart home-based readers for the early detection of health problems.