Ersin Bayram

Genetic algorithms and self-organizing maps: a powerful combination for modeling complex QSAR and QSPR problems

Modeling non-linear descriptor-target activity/property relationships with many dependent descriptors has been a long-standing challenge in the design of biologically active molecules. In an effort to address this problem, we couple the supervised self-organizing map with the genetic algorithm. Although self-organizing maps are non-linear and topology-preserving techniques that hold great potential for modeling and decoding relationships, the large number of descriptors in typical quantitative structure--activity relationship or quantitative structure--property relationship analysis may lead to spurious correlation(s) and/or difficulty in the interpretation of resulting models. To reduce the number of descriptors to a manageable size, we chose the genetic algorithm for descriptor selection because of its flexibility and efficiency in solving complex problems. Feasibility studies were conducted using six different datasets, of moderate-to-large size and moderate-to-great diversity; each with a different biological endpoint. Since favorable training set statistics do not necessarily indicate a highly predictive model, the quality of all models was confirmed by withholding a portion of each dataset for external validation. We also address the variability introduced onto modeling through dataset partitioning and through the stochastic nature of the combined genetic algorithm supervised self-organizing map method using the z-score and other tests. Experiments show that the combined method provides comparable accuracy to the supervised self-organizing map alone, but using significantly fewer descriptors in the models generated. We observed consistently better results than partial least squares models. We conclude that the combination of genetic algorithms with the supervised self-organizing map shows great potential as a quantitative structure--activity/property relationship modeling tool.

Supervised self-organizing maps in drug discovery. 2. Improvements in descriptor selection and model validation

The modeling of nonlinear descriptor-target relationships is a topic of considerable interest in drug discovery. We, herein, continue reporting the use of the self-organizing map-a nonlinear, topology-preserving pattern recognition technique that exhibits considerable promise in modeling and decoding these relationships. Since simulated annealing is an efficient tool for solving optimization problems, we combined the supervised self-organizing map with simulated annealing to build high-quality, highly predictive quantitative structure-activity/property relationship models. This technique was applied to six data sets representing a variety of biological endpoints. Since a high statistical correlation in the training set does not indicate a highly predictive model, the quality of all the models was confirmed by withholding a portion of each data set for external validation. Finally, we introduce new cross-validation and dynamic partitioning techniques to address model overfitting and assessment.

Minimum reliable scale selection in 3D

Multiscale analysis is often required in image processing applications because image features are optimally detected at different levels of resolution. With the advance of high-resolution 3D imaging, the extension of multiscale analysis to higher dimensions is necessary. This paper extends an existing 2D scale selection method, known as the minimum reliable scale, to 3D volumetric images. The method is applied to 3D boundary detection and is illustrated in examples from biomedical imaging. The experimental results show that the 3D scale selection improves the detection of edges over single scale operators using as few as three different scales.

Confidence based anisotropic filtering of magnetic resonance images

Image filtering is an important off-line image processing technique to improve the signal-to-noise ratio (SNR) and/or contrast-to-noise ratio (CNR) of acquired images. The major drawback of filtering is that it often blurs the fine structures and object boundaries in the image along with noise. Anisotropic diffusive filtering techniques incorporate gradient information to blur homogeneous regions while preserving the boundaries and interesting structures. Unfortunately, their performance is limited in low contrast regions and around fuzzy boundaries. This paper introduces a multi-scale confidence based conductance function to address the limitations of anisotropic diffusive filtering. Experiments on phantom and magnetic resonance (MR) images have been performed using both our method and the gradient-based anisotropic diffusive filtering for comparison purposes.Wiener filter restoration, followed by a difference operator, is used to estimate the standard deviation of the noise based on the additive noise assumption. Simulation studies show that a 5 /spl times/ 5 Wiener filter gives an estimate of noise within a 5% error margin. A careful examination of the conductance map in the brain MR image reveals that a wide band of zero conductance region is seen around blurred boundaries. To blend these regions without allowing a generous blurring, a small constant can be added to the conductance function. A better approach will be incorporating the second derivative information into the conductance function. As edges are defined at the zero-crossings of the second derivative response, the strength of the second derivative response can be used as a measure of distance to a boundary. Unfortunately, in the discrete domain, edges generally fall off pixel locations. Thus, second derivative strength would not be a quite reliable measure, unless interpolation and subsampling are employed.

Confidence-based anisotropic filtering of magnetic resonance images

Wiener filter restoration, followed by a difference operator, is used to estimate the standard deviation of the noise based on the additive noise assumption. Simulation studies show that a 5 /spl times/ 5 Wiener filter gives an estimate of noise within a 5% error margin. A careful examination of the conductance map in the brain MR image reveals that a wide band of zero conductance region is seen around blurred boundaries. To blend these regions without allowing a generous blurring, a small constant can be added to the conductance function. A better approach will be incorporating the second derivative information into the conductance function. As edges are defined at the zero-crossings of the second derivative response, the strength of the second derivative response can be used as a measure of distance to a boundary. Unfortunately, in the discrete domain, edges generally fall off pixel locations. Thus, second derivative strength would not be a quite reliable measure, unless interpolation and subsampling are employed.

Automatic scale selection for medical image segmentation

The scale of interesting structures in medical images is space variant because of partial volume effects, spatial dependence of resolution in many imaging modalities, and differences in tissue properties. Existing segmentation methods either apply a single scale to the entire image or try fine-to-coarse/coarse-to-fine tracking of structures over multiple scales. While single scale approaches fail to fully recover the perceptually important structures, multi-scale methods have problems in providing reliable means to select proper scales and integrating information over multiple scales. A recent approach proposed by Elder and Zucker addresses the scale selection problem by computing a minimal reliable scale for each image pixel. The basic premise of this approach is that, while the scale of structures within an image vary spatially, the imaging system is fixed. Hence, sensor noise statistics can be calculated. Based on a model of edges to be detected, and operators to be used for detection, one can locally compute a unique minimal reliable scale at which the likelihood of error due to sensor noise is less than or equal to a predetermined threshold. In this paper, we improve the segmentation method based on the minimal reliable scale selection and evaluate its effectiveness with both simulated and actual medical data.

MR tagging from a signal processing perspective

Although magnetic resonance (MR) tagging has been shown to be a useful tool in myocardial motion quantification, its clinical utilization is limited as current available methods generally either lack computational speed or require extensive user intervention. Recently, the harmonic phase imaging (HARP) technique has been proposed to look at the phase information of the tagged images. HARP imaging promises to overcome the limitations of existing methods in terms of both computational speed and automation. Motivated by this work, we present mathematical analysis providing a signal processing perspective on the HARP technique. This new perspective provides a clearer understanding of how tags can be accurately tracked using highly-filtered data.

Spatial-spectral fat suppression in phase-contrast coronary flow imaging

Magnetic resonance (MR) phase contrast (PC) imaging holds great promise as a noninvasive diagnostic tool for coronary heart disease (CHD) by measuring the blood flow in coronary arteries. Fat signal from the vasculature bed and chest wall generates artifacts in terms of motion ghosts, shifting, and blurring; which hinders vessel segmentation (needed for flow analysis). For a 1.5 Tesla system, the resonance frequency of fat is approximately 220 Hz slower than water. Spatial-spectral (SPSP) pulses utilize this resonance frequency difference to selectively excite either fat or water at a specific location. While SPSP pulses have been shown to be superior over conventional fat saturation pulses, their long durations (10-15 ms) have hindered their use in MR coronary flow imaging, which requires high temporal resolution. Our goal is to design short duration SPSP pulses suitable for PC coronary flow imaging.