Artur Chodorowski

Active contour models: application to oral lesion detection in color images

This paper presents the application of active contour models (Snakes) for the segmentation of oral lesions in medical color images acquired from the visual part of the light spectrum. The aim is to assist the clinical expert in locating potentially cancerous cases for further analysis (e.g. classification of cancerous vs. non-cancerous lesions). In order to apply the conventional Snake formulation, color images were converted into single-band images. A number of different single-bands were evaluated including those resulting from the original and normalized RGB, perceptual HSI space, I/sub 1/I/sub 2/I/sub 3/, and the Fisher discriminant function. Examples of Snake segmentation results of oral lesions are presented.

Color Lesion Boundary Detection Using Live Wire

The boundaries of oral lesions in color images were detected using a live-wire method and compared to expert delineations. Multiple cost terms were analyzed for their inclusion in the final total cost function including color gradient magnitude, color gradient direction, Canny edge detection, and Laplacian zero crossing. The gradient magnitude and direction cost terms were implemented so that they acted directly on the three components of the color image, instead of using a single derived color band. The live-wire program was shown to be considerably more accurate and faster compared to manual segmentations by untrained users.

A novel Bayesian approach to adaptive mean shift segmentation of brain images

We present a novel adaptive mean shift (AMS) algorithm for the segmentation of tissues in magnetic resonance (MR) brain images. In particular we introduce a novel Bayesian approach for the estimation of the adaptive kernel bandwidth and investigate its impact on segmentation accuracy. We studied the three class problem where the brain tissues are segmented into white matter, gray matter and cerebrospinal fluid. The segmentation experiments were performed on both multi-modal simulated and real patient TI-weighted MR volumes with different noise characteristics and spatial inhomogeneities. The performance of the algorithm was evaluated relative to several competing methods using real and synthetic data. Our results demonstrate the efficacy of the proposed algorithm and that it can outperform competing methods, especially when the noise and spatial intensity inhomogeneities are high.

Oral lesion classification using true-color images

The aim of the study was to investigate effective image analysis methods for the discrimination of two oral lesions, oral lichenoid reactions and oral leukoplakia, using only color information. Five different color representations (RGB, Irg, HSI, I1I2I3 and La*b*) were studied and their use for color analysis of mucosal images evaluated. Four common classifiers (Fishers linear discriminant, Gaussian quadratic, kNN-Nearest Neighbor and Multilayer Perceptron) were chosen for the evaluation of classification performance. The feature vector consisted of the mean color difference between abnormal and normal regions extracted from digital color images. Classification accuracy was estimated using resubstitution and 5-fold crossvalidation methods. The best classification results were achieved in HSI color system and using linear discriminant function. In total, 70 out of 74 (94.6%) lichenoid reactions and 14 out of 20 (70.0%) of leukoplakia were correctly classified using only color information.

Non-invasive EEG source localization using particle swarm optimization: A clinical experiment

One of the most important steps of pre-surgical diagnosis in patients with medically intractable epilepsy is to find the precise location of the epileptogenic foci. An Electroencephalography (EEG) is a non-invasive standard tool used at epilepsy surgery center for pre-surgical diagnosis. In this paper a modified particle swarm optimization (MPSO) method is applied to a real EEG data, i.e., a somatosensory evoked potentials (SEPs) measured from a healthy subject, to solve the EEG source localization problem. A high resolution 1 mm hexahedra finite element volume conductor model of the subjects head was generated using T1-weighted magnetic resonance imaging data. An exhaustive search pattern and the MPSO method were then applied to the peak of the averaged SEPs data. The non-invasive EEG source analysis methods localized the somatosensory cortex area where our clinical expert expected the received SEPs. The proposed inverse problem solver found the global minima with acceptable accuracy and reasonable number of iterations.

Computer analysis in oral lichenoid reactions

To improve diagnostic procedures and facilitate clinical decision-making, computer-assisted image analysis was performed on color slides from 30 patients with histopathologically verified oral lichenoid reactions. Areas from white hyperkeratotic and adjacent red inflamed areas of the lesions were selected and subjected to image analysis. The digitization of the color slides was done by means of an image scanner, and the digital information was transmitted to a personal computer for subsequent feature extraction and analysis. The different oral lesions were characterized as the difference in mean values between white hyperkeratotic and red inflamed areas, respectively, compared with clinically normal tissue. Statistical analyses were made on three different color systems: Red-Green-Blue (RGB), normalized red-green-blue (rgb), and Intensity-Hue-Saturation (IHS). The results showed statistically significant differences in all color systems for both the hyperkeratotic areas and adjacent inflammatory reactions. A linear correlation was obtained when the results of the image analysis of color variations were compared with a clinical score system for hyperkeratosis and inflammation evaluated by two investigators independently.

An attempt to estimate out-of-plane lung nodule elongation in tomosynthesis images

In chest tomosynthesis (TS) the most commonly used reconstruction methods are based on Filtered Back Projection (FBP) algorithms. Due to the limited angular range of x-ray projections, FBP reconstructed data is typically associated with a low spatial resolution in the out-of-plane dimension. Lung nodule measures that depend on depth information such as 3D shape and volume are therefore difficult to estimate. In this paper the relation between features from FBP reconstructed lung nodules and the true out-of-plane nodule elongation is investigated and a method for estimating the out-of-plane nodule elongation is proposed. In order to study these relations a number of steps that include simulation of spheroidal-shaped nodules, insertion into synthetic data volumes, construction of TS-projections and FBP-reconstruction were performed. In addition, the same procedure was used to simulate nodules and insert them into clinical chest TS projection data. The reconstructed nodule data was then investigated with respect to in-plane diameter, out-of-plane elongation, and attenuation coefficient. It was found that the voxel value in each nodule increased linearly with nodule elongation, for nodules with a constant attenuation coefficient. Similarly, the voxel value increased linearly with in-plane diameter. These observations indicate the possibility to predict the nodule elongation from the reconstructed voxel intensity values. Such a method would represent a quantitative approach to chest tomosynthesis that may be useful in future work on volume and growth rate estimation of lung nodules.

A fully automatic unsupervised segmentation framework for the brain tissues in MR images

This paper presents a novel fully automatic unsupervised framework for the segmentation of brain tissues in magnetic resonance (MR) images. The framework is a combination of our proposed Bayesian-based adaptive mean shift (BAMS), a priori spatial tissue probability maps and fuzzy c-means. BAMS is applied to cluster the tissues in the joint spatialintensity feature space and then a fuzzy c-means algorithm is employed with initialization by a priori spatial tissue probability maps to assign the clusters into three tissue types; white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF). The proposed framework is validated on multimodal synthetic as well as on real T1-weighted MR data with varying noise characteristics and spatial intensity inhomogeneity. The performance of the proposed framework is evaluated relative to our previous method BAMS and other existing adaptive mean shift framework. Both of these are based on the mode pruning and voxel weighted k-means algorithm for classifying the clusters into WM, GM and CSF tissue. The experimental results demonstrate the robustness of the proposed framework to noise and spatial intensity inhomogeneity, and that it exhibits a higher degree of segmentation accuracy in segmenting both synthetic and real MR data compared to competing methods.

Support vector machines for oral lesion classification

We investigate support vector machines (SVM) in the context of oral lesion classification using digital color images as input. Two common lesions of similar visual appearance to the human observer were evaluated: oral leukoplakia, which is a potentially pre-cancerous lesion, and oral lichenoid reactions (with subclasses of atrophic, plaqueformed and reticular reactions), which are usually harmless lesions. In total, 89% (212 out of 238, 5-fold CV) were correctly classified in a two-class problem (precancerous vs. non-pre-cancerous) and 78% (61 out of 78, hold-out) into four classes (complete classification). The proposed method can be used as a decision support tool in CADx systems for oral lesion classification and detection of potentially pre-cancerous lesions.

Automated Estimation of In-plane Nodule Shape in Chest Tomosynthesis Images

The purpose of this study was to develop an automated segmentation method for lung nodules in chest tomosynthesis images. A number of simulated nodules of different sizes and shapes were created and inserted in two different locations into clinical chest tomosynthesis projections. The tomosynthesis volumes were then reconstructed using standard cone beam filtered back projection, with 1 mm slice interval. For the in-plane segmentation, the central plane of each nodule was selected. The segmentation method was formulated as an optimization problem where the nodule boundary corresponds to the minimum of the cost function, which is found by dynamic programming. The cost function was composed of terms related to pixel intensities, edge strength, edge direction and a smoothness constraint. The segmentation results were evaluated using an overlap measure (Dice index) of nodule regions and a distance measure (Hausdorff distance) between true and segmented nodule. On clinical images, the nodule segmentation method achieved a mean Dice index of 0.96 ± 0.01, and a mean Hausdorff distance of 0.5 ± 0.2 mm for isolated nodules and for nodules close to other lung structures a mean Dice index of 0.95 ± 0.02 and a mean Hausdorff distance of 0.5 ± 0.2 mm. The method achieved an acceptable accuracy and may be useful for area estimation of lung nodules.