Siegfried J. Pöppl

Feature selection for optimized skin tumor recognition using genetic algorithms

In this paper, a new approach to computer supported diagnosis of skin tumors in dermatology is presented. High resolution skin surface profiles are analyzed to recognize malignant melanomas and nevocytic nevi (moles), automatically. In the first step, several types of features are extracted by 2D image analysis methods characterizing the structure of skin surface profiles: texture features based on cooccurrence matrices, Fourier features and fractal features. Then, feature selection algorithms are applied to determine suitable feature subsets for the recognition process. Feature selection is described as an optimization problem and several approaches including heuristic strategies, greedy and genetic algorithms are compared. As quality measure for feature subsets, the classification rate of the nearest neighbor classifier computed with the leaving-one-out method is used. Genetic algorithms show the best results. Finally, neural networks with error back-propagation as learning paradigm are trained using the selected feature sets. Different network topologies, learning parameters and pruning algorithms are investigated to optimize the classification performance of the neural classifiers. With the optimized recognition system a classification performance of 97.7% is achieved.

Atlas-based segmentation of bone structures to support the virtual planning of hip operations

Two 3-D digitised atlases of a female and a male pelvis were generated to support the virtual 3-D planning of hip operations. The anatomical atlases were designed to replace the interactive, time-consuming pre-processing steps for the virtual operation planning. Each atlas consists of a labelled reference CT data set and a set of anatomical point landmarks. The paper presents methods for the automatic transfer of these anatomical labels to an individual patient data set. The labelled patient data are used to generate 3-D models of the patients bone structures. Besides the anatomical labelling, the determination of measures, like angles, distances or sizes of contact areas, is important for the planning of hip operations. Thus, algorithms for the automatic computation of orthopaedic parameters were implemented. A first evaluation of the presented atlas-based segmentation method shows a correct labelling of 98.5% of the bony voxels.

Virtual planning of hip operations and individual adaption of endoprostheses in orthopaedic surgery

The introduction of virtual reality techniques in medicine opens up new possibilities for the planning of interventions. The presented software system for virtual operation planning in orthopaedic surgery (VIRTOPS) enables the virtual preoperative 3D planning and simulation of pelvis and hip operations. It is used to plan operations of bone tumours with endoprosthetic reconstruction of the hip based on multimodal image information. The operation and the endosprothetic reconstruction of the pelvis are simulated using virtual reality techniques. Stereoscopic visualisation techniques and 3D input devices support the 3D interaction with the virtual 3D models. The main task of the preoperative planning process is the individual design of an anatomically adaptable modular prosthesis. The placement and the design of the endoprosthesis are supported by different functions and visualisation techniques. The resulting 3D images and movies can be used for the documentation of the operation planning procedure, as well as, for the preoperative information of the patient.

ACMD: A Practical Tool for Automatic Neural Net Based Learning

Although neural networks have many appealing properties, yet there is neither a systematic way how to set up the topology of a neural network nor how to determine its various learning parameters. Thus an expert is needed for fine tuning. If neural network applications should not be realisable only for publications but in real life, fine tuning must become unnecessary. We developed a tool called ACMD (Approximation and Classification of Medical Data) that is demonstrated to fulfil this demand. Moreover referring to six medical classification and approximation problems of the PROBEN1 benchmark collection this approach will be shown even to outperform fine tuned networks.

A new neural network approach classifies olfactory signals with high accuracy

Artificial neural networks (ANN) become more significant in signal processing. Because ANN still have some drawbacks we developed a new neural network tool named ACMD considering several methods of resolution, existing ones as well as new ones. In order to demonstrate the capabilities of ACMD in the field of food quality, we classified signals from an electronic nose smelling different types of edible oil and honey. The accuracies achieved by ACMD were evidently greater than the accuracies obtained by ANN trained by other well-known methods. As a conclusion it seems to be worthwhile considering sophisticated ANN strategies like those integrated in ACMD.

The capabilities of artificial neural networks in body composition research

Abstract.When estimating in vivo body composition or combining such estimates with other results, multiple variables must be taken into account (e. g. binary attributes such as gender or continuous attributes such as most biosignals). Standard statistical models, such as logistic regression and multivariate analysis, presume well-defined distributions (e. g. normal distribution); they also presume independence among all inputs and only linear relationships, yet rarely are these requirements met in real life. As an alternative to these models, artificial neural networks can be used. In the present work, we describe the pre-processing and multivariate analysis of data using neural network techniques, providing examples from the medical field and making comparisons with classic statistical approaches. We also address the criticisms raised regarding neural network techniques and discuss their potential improvement.

Atlas-Based Recognition of Anatomical Structures and Landmarks to Support the Virtual Three-Dimensional Planning of Hip Operations

This paper describes methods for the atlas-based segmentation of bone structures of the hip, the automatic detection of anatomical point landmarks and the computation of orthopedic parameters. An anatomical atlas was designed to replace interactive, time-consuming pre-processing steps needed for the virtual planning of hip operations. Furthermore, a non-linear gray value registration of CT data is used to recognize different bone structures of the hip. A surface based registration algorithm enables the robust and precise detection of anatomical point landmarks. Furthermore the determination of quantitative parameters, like angles, distances or sizes of contact areas, is important for the planning of hip operations. Based on segmented bone structures and detected landmarks algorithms for the automatic computation of orthopedic parameters were implemented. A first evaluation of the presented methods will be given at the end of the paper.

Automatic Classsification of Skin Tumours with High Resolution Surface Profiles

This paper describes a new approach to automatic classification of melanocytic tumours based on features extracted from profilometric data. The clinical accuracy of dermatologists in identifying these tumours is only approximately 75%. Automatic classification is based on high resolution skin surface profiles of 4×4 mm2 size with 125 sample points per mm, generated with a laser profilometer. Three categories of profile features are extracted: Textural features, Fourier features and fractal features. Feature selection is performed to determine an optimal feature subset. As a quality measure for a given feature subset, the error rate of the nearest neighbour classifier estimated with the leaving-one-out method is used. With the optimal feature subset, feed forward neural networks with error backpropagation as learning function are trained. Several neural networks with different network topologies and learning parameters were trained to compare the classification performance. A three layer network with one hidden layer consisting of 20 units has shown the best performance of all considered neural networks with a classification error rate of 13.4%. The best results using the nearest neighbour classifier achieved an error rate of 6.8%.

Characterisation and Classification of Brain Tumours in Three-Dimensional MR Image Sequences

In this paper a new approach for quantitative description and recognition of brain tumours in three-dimensional MR image sequences is presented. In radiological diagnostics, tumour features like shape, irregularity of tumour borders, contrast-enhancement etc. are used in a qualitative way. The presented image analysis methods extract quantitative descriptions of two- and three-dimensional tumour features. The tumour shape is described by ellipsoid approximations and the analysis of tumour profiles. Fractal features quantify irregularities and self-similarities of tumour contours. The internal tumour structure and especially the homogeneity of tumour tissue is analysed by means of texture analysis. The extracted features reflect 3D information about tumour morphology. The high-dimensional feature information is analyzed by the nearest neighbour classifier. Furthermore, in combination with the leaving-one-out method the nearest neighbour classifier is used to select proper feature subsets discriminating different tumours types. In a clinical study, the developed methods were applied to image sequences with the four most frequent brain tumours: meningiomas, astrocytomas, glioblastomas, and metastases. The clinical accuracy of radiologists in identifying brain tumours is approx. 80% [KWG+89], The automatic recognition of tumour types achieves a classification rate of 93%.

Artificial neural network analysis: a novel application for predicting site-specific bone mineral density

Abstract.Dual X-ray absorptiometry (DXA), which is the most commonly used method for the diagnosis and followup of human bone health, is known to produce accurate estimates of bone mineral density (BMD). However, high costs and problems with availability may prevent its use for mass screening. The objective of the present study was to estimate BMD values for healthy persons and those with conditions known to be associated with BMD, using artificial neural networks (ANN). An ANN was used to quantitatively estimate site-specific BMD values in comparison with reference values obtained by DXA (i. e. BMDspine, BMDpelvis, and BMDtotal). Anthropometric measurements (i. e. sex, age, weight, height, body mass index, waist-to-hip ratio, and the sum of four skinfold thicknesses) were fed to the ANN as independent input variables. The estimates based on four input variables were generated as output and were generally identical to the reference values for all studied groups. We believe the ANN is a promising approach for estimating and predicting site-specific BMD values using simple anthropometric measurements.