Main page


Preface

Program
summary

Research
projects
1994-1998

Research
projects
1998-1999

Product
development
projects
1994-1999

6. Applications in the forest industry and in medical informatics (IIS:COMPSOFT + NESUMED II)

6.1 Intelligent information systems: computing methods and software developments (IIS:COMPSOFT)

Abstract

The forest products industry has in recent years found itself in a situation where information of various kinds and extents is becoming more easily available and transferable. Enhancements of computer systems and technical platforms both for process control and information management purposes has resulted in information becoming instantly available for refinement and knowledge acquisition tasks. In this paper we describe the possibilities of data analysis and knowledge acquisition for the development of web break risk indication. Software has been developed and integrated into a real process.

Results

The IIS:COMPSOFT project has provided data analysis and software developments in web break indidication for paper machines. The Break Indicator program computes break risk values based on signals obtained from a DDE server. The Break Indicator can connect directly to the DDE server of the process monitoring software. The Break Indicator can also connect to a buffering file which in turn is connected to the process monitoring software. A buffering file can be arranged e.g. using Microsoft Excel. DDE addressing information is read from a data file.

Figure 1. The break indicator (connected to an idle process).

We found it important not to include functionality already available in the process monitoring systems, but rather to complement the monitoring system with the additional functionality of providing break risk predictions. The software is not technically a plug-in to the monitoring systems but was designed to act as such. The software runs under Windows NT and was developed as a Visual Basic application. The programming environment was selected so as to simplify maintenance, and enable addition of further functionality without having to cope with the complexity of software development environment. Communication with the data server uses DDE (Dynamic Data Exchange), either with the break indicator directly connected to the DDE server, or connected via a buffering node, e.g. a MS Excel file. Optionally, a user may communicate via netDDE, and thus the break indicator can be used also in a mobile extension (see Figure 2).

Figure 2. A DDE server can be connected internally to the break indicator or via a network at a particular IP address. A mobile phone connection establishing an IP connection makes the DDE server similarly reachable.

Project information

Participants

  • Åbo Akademi University
  • Valmet Oyj (Jyväskylä)
  • UPM (Jämsänkoski)

Project dates

March 1, 1998 – December 31, 1999

Project volume

FIM 590.000, 21 man months

Project manager

Prof. Patrik Eklund
Umeå University, Department of Computing Science
SE-901 87 Umeå, Sweden
Phone: +46-90-7869914
E-mail: peklund@cs.umu.se

Publications

[1]  P. Eklund, F. Georgsson, Unravelling the thrill of metric image spaces, Discrete Geometry for Computer Imagery (eds. G. Bertrand, M. Couprie, L. Perroton), Lecture Notes in Computer Science 1568, Springer Verlag, 1999, pp. 275-285.

[2]  P. Eklund, L. Kallin, G. Selén, Computational intelligence for medical information analysis and refinement, Applications of Fuzzy \& Neuro-Fuzzy Systems in Medicine and Bio-Medical Engineering (eds. H .N. Teodorescu, A. Kandel, L.C. Jain), CRC Press, 1999.

[3]  P. Eklund, T. Riissanen, Transformation functions and network parameters for the break indicator, Internal report, Åbo Akademi University, August 14, 1998, revised November 7, 1999.

[4]  P. Eklund, T. Riissanen, Forest informatics, IFORS-SPC-9, April 25-27, 1999, Turku, Finland, pp. 26-29.

[5]  P. Eklund, T. Riissanen, Software development for web break risk indication, 2000 TAPPI PCE&I, March 26-30, 2000, Williamsburg Lodge, Williamsburg VA.

[6]  P. Eklund, J. Zhou, Comparison of learning strategies for parameter identification in rule based systems, J. Fuzzy Sets and Systems, 106 (1999), pp. 321-333.

[7]  M. Persson, J. Bohlin, P. Eklund, Development and maintenance of guideline-based decision support for pharmacological treatment of hypertension, Computer Methods and Programs in Medicine, to appear.

[8]  J. Karlsson, P. Eklund, C.-G. Hallgren, J. Sjödin, Data warehousing as a basis for web-based documentation of data mining and analysis, Medical Informatics Europe,'99 (eds. P. Kokol, B. Zupan, J. Stare, M. Premik, R. Engelbrecht), IOS Press, 1999, pp. 423-427.


6.2 Neural networks and fuzzy logic in medical informatics (NESUMED II)

Abstract

We have applied neural networks to calculate the risk of Down's syndrome from biochemical screening data of the mothers. In the preliminary tests with neural networks, we have been able to obtain the same accuracy as  with the statistical formula based on the multivariate Gaussian model, which is currently used in most European screening centres. Artificial neural networks (ANN) are able to learn non-linear models from example cases. The use of ANNs can be desirable when the number of example cases is large and the model that will be trained includes several influencing factors. Many patient-specific factors influence the levels of AFP and hCG and these effects can be learned from data obtained from normal pregnancies. Furthermore, the effects are most likely to be non-linear and thus hard to model using conventional statistical techniques.  Consequently, ANNs should be very suitable for this task.  In the NESUMED II project we compared an ANN to the weight adjustment formula routinely used in risk calculation software and demonstrated how the advantages of ANNs can be included in the risk calculation for Down syndrome. Only gestational age, maternal weight and maternal age were used as variables.

Screening for prostate cancer using PSA and free PSA measurements is widely used in developed countries. The accuracy of screening can be enhanced by linking other information about the patient to the model. For this purpose we have compared ANNs, fuzzy logic and a neuro-fuzzy tool with logistic regression models to see whether they are useful in tackling the problem.

Trabecular bone architecture adapts to mechanical loading, which results in functional groups of trabeculae that have a common or slowly changing orientation and almost regular spacing. Our method efficiently utilizes this knowledge. The method builds a pseudo-reconstruction of the trabecular bone based on the relatively clear radiographic depiction of trabecular structures.

Recognition of trabecular elements in radiographs is a complicated task, mainly because three-dimensional structure is projected onto two-dimensional radiographs. However, in a suitably aligned radiograph, the trabecular bone is seen as a network of fuzzy bars, which are usually organized in two roughly orthogonal main orientations making the extraction of the trabecules possible. Based on this structure, the method builds a pseudo-reconstruction of the trabecular structure. This step uses sophisticated image analysis algorithms to detect individual bone struts in various trajectorial groups. The reconstruction is used to derive different types of microstructure descriptors, such as thickness and spacing of trabecular struts, orientation distribution, anisotropy and connectivity. Because of the approach of using a pseudo-reconstruction of the structure as an intermediate step, there is a direct correspondence with the measured descriptors and real structure of the bone. This is useful because in many bone disorders the normal architecture changes in a systematic and predictable way. Using this kind application-specific knowledge, very subtle changes can be detected. Furthermore, it is becomes possible to measure changes in individual elements of the trabecular bone, which can be used in follow-up applications.

The key techniques used were

  1. Gabor wavelet based directional filtering to suppress trabecules having an orientation outside the current scope.
  2. Detection of possible trabecules and subsequent validation using application-specific algorithms.
  3. Accurate segmentation of trabecules using deformable contours.
  4. Building of the pseudo-reconstruction of the trabecular group using application-specific ‘tailor-made’ algorithms.

Results

For Down’s syndrome screening we were able to show that:

  • Using a neural network (multilayer perceptron) we could achieve the same diagnostic accuracy as with statistical tools.
  • The model differed among different populations and, therefore, the model has to be calculated using data from the local population.
  • When fitting the model to another population, additional learning using neural networks turned out to be problematic, since the number of examples in the local population was too small. Therefore, better performance in transferring a model to another population was obtained when using traditional statistical parametric tools, where we could use population characteristics such as mean and standard deviation in adjusting the formula to local needs.
In screening for prostate cancer, combining laboratory medicine with background data from patients and physical examination, we were able to show that:

  • Logistic regression turned out to be at least as good as a multilayer perceptron in the classification task.
  • Some preliminary data showed that with neuro-fuzzy tools we may be able to develop a model that is better than that based on logistic regression, but more data is needed for comparison.
In the osteoporosis project we developed new algorithms which have not been reported. A prototype software for radiographic trabecule bone analysis has almost been finished. Data collection from 800 elderly patients has been completed. The project will continue in 2000 and medical results are expected within the next six months.

Project information

Participants

  • Medical Informatics Research Centre in Turku (MIRCIT), University of Turku
  • Department of Computer Science, University of Turku
  • Härkätie Health Care Centre (Lieto)
  • Wallac Oy (Turku)
  • Leiras Oy (Turku)

Project dates

March 1, 1998 – December 31, 1999

Project volume

FIM 1.150.000, 40 man months

Project manager

MD Jari Forsström

Medical Informatics Research Centre in Turku, MIRCIT
University of Turku
Kiinamyllynkatu 4-8, 20520 Turku, Finland
Phone: +358-2-2612 914
Fax: +358-2-2613 920
E-mail: jari.forsstrom@utu.fi

Publications

[1] Selén G, Forsström J. Local adaptation of artificial neural networks in detecting Down's syndrome. Scand J Clin Lab Invest 1998; 58 (Suppl 228):101.

[2] Virtanen A, Gomari M, Kranse R, Stenman U-H. Estimation of prostate cancer probability by logistic regression: free and total prostate-specific antigen, digital rectal examination, and heredity are significant variables. Clin Chem 1999; 45:987-994.

[3]  Forsström J, Räty R, Salonen R, Stenman U-H, Helenius H, Virtanen A, Selén G. Down syndrome screening: using artificial neural networks to model patient-specific factors on biochemical markers. Scand J Clin Lab Invest (in press).

WEB-sites

NESUMED: http://users.utu.fi/jarifors/nesumed.html (in English)

NESUMED II: http://www2.cs.utu.fi/staff/jaakko.jarvi@nesumed (in Finnish)

MIRCIT: http://www.utu.fi/research/mircit (Partly in English)

Pattern recognition in osteoporosis: http://www2.cs.utu.fi/research/projects/analysis.medi.html



jukka.iivarinen@hut.fi
http://www.cis.hut.fi/neuronet/Tekes/6.shtml
Wednesday, 29-Nov-2000 10:24:46 EET