Juha Koikkalainen

Methods of Artificial Enlargement of the Training Set for Statistical Shape Models

Due to the small size of training sets, statistical shape models often over-constrain the deformation in medical image segmentation. Hence, artificial enlargement of the training set has been proposed as a solution for the problem to increase the flexibility of the models. In this paper, different methods were evaluated to artificially enlarge a training set. Furthermore, the objectives were to study the effects of the size of the training set, to estimate the optimal number of deformation modes, to study the effects of different error sources, and to compare different deformation methods. The study was performed for a cardiac shape model consisting of ventricles, atria, and epicardium, and built from magnetic resonance (MR) volume images of 25 subjects. Both shape modeling and image segmentation accuracies were studied. The objectives were reached by utilizing different training sets and datasets, and two deformation methods. The evaluation proved that artificial enlargement of the training set improves both the modeling and segmentation accuracy. All but one enlargement techniques gave statistically significantly (p < 0.05) better segmentation results than the standard method without enlargement. The two best enlargement techniques were the nonrigid movement technique and the technique that combines principal component analysis (PCA) and finite element model (FEM). The optimal number of deformation modes was found to be near 100 modes in our application. The active shape model segmentation gave better segmentation accuracy than the one based on the simulated annealing optimization of the model weights.

Reconstruction of 3-D head geometry from digitized point sets: an evaluation study

In this paper, we evaluate different methods to estimate patient-specific scalp, skull, and brain surfaces from a set of digitized points from the targets scalp surface. The reconstruction problem is treated as a registration problem: An a priori surface model, consisting of the scalp, skull, and brain surfaces, is registered to the digitized surface points. The surface model is generated from segmented magnetic resonance (MR) volume images. We study both affine and free-form deformation (FFD) registration, the use of average models, the averaging of individual registration results, a model selection procedure, and statistical deformation models. The registration algorithms are mainly previously published, and the objective of this paper is to evaluate these methods in this particular application with sparse data. The main interest of this paper is to generate geometric head models for biomedical applications, such as electroencephalography and magnetoencephalographic. However, the methods can also be applied to other anatomical regions and to other application areas. The methods were validated using 15 MR volume images, from which the scalp, skull, and brain were manually segmented. The best results were achieved by averaging the results of the FFD registrations of the database: the mean distance from the manually segmented target surface to a deformed a priori model surface for the studied anatomical objects was 1.68-2.08 mm, depending on the point set used. The results support the use of the evaluated methods for the reconstruction of geometric models in applications with sparse data.

Artificial enlargement of a training set for statistical shape models: application to cardiac images

Different methods were evaluated to enlarge artificially a training set which is used to build a statistical shape model. In this work, the shape model was built from MR data of 25 subjects and it consisted of ventricles, atria and epicardium. The method adding smooth non-rigid deformations to original training set examples produced the best results. The results indicated also that artificial deformation modes model better an unseen object than an equal number of standard PCA modes generated from original data.

Artificially enlarged training set in image segmentation

Due to small training sets, statistical shape models constrain often too much the deformation in medical image segmentation. Hence, an artificial enlargement of the training set has been proposed as a solution for the problem. In this paper, the error sources in the statistical shape model based segmentation were analyzed and the optimization processes were improved. The method was evaluated with 3D cardiac MR volume data. The enlargement method based on non-rigid movement produced good results--with 250 artificial modes, the average error for four-chamber model was 2.11 mm when evaluated using 25 subjects.

Image segmentation with the combination of the PCA- and ICA-based modes of shape variation

How to constrain the deformations in deformable model-based image segmentation is a well-studied issue. Many methods that use the modes of shape variation generated from a training set shapes have been introduced. Most of these methods rely on principle component analysis (PCA) to statistically model the variability in the training set. Independent component analysis (ICA) has been proposed for this purpose, too. In this paper, we combine the PCA- and ICA-based modes of shape variation using a consecutive approach: an a priori model is deformed first by the PCA modes, which represent the global shape variability in the training set, and then, by the ICA modes, which have a more local character. The method is validated using a set of three-dimensional (3D) brain MR images. The results prove that by applying the ICA modes after the PCA modes the accuracy of image segmentation is statistically significantly (p < 0.05) improved.

Four-Chamber 3-D Statistical Shape Model from Cardiac Short-Axis and Long-Axis MR Images

We describe a new statistical atlas of the heart consisting of atrias, ventricles and epicardium. The atlas was constructed by combining information on standard short-axis and long-axis cardiac MR images. The variability of the shape was modeled in the atlas by a statistical deformation model and by non-parametric probability distributions. The atlas has been built from 16 subjects.

Segmentation of Cardiac Structures Simultaneously from Short- and Long-Axis MR Images

We introduce a framework for the automatic segmentation of the ventricles, atria, and epicardium simultaneously from cardiac magnetic resonance (MR) volumes. The basic idea is to utilize both short-axis (SA) and long-axis (LA) MR volumes. Consequently, anatomical information is available from the whole heart volume. In this paper, the framework is used with deformable model based registration and segmentation methods to segment the cardiac structures. A database consisting of the cardiac MR volumes of 25 healthy subjects is used to validate the methods.

Model Library for Deformable Model-Based Segmentation of 3-D Brain MR-Images

A novel method to use model libraries in segmentation is introduced. Using similarity measures one model from a model library is selected. This model is then used in model-based segmentation. The proposed method is simple, straightforward and fast. Various similarity measures, both voxel and edge measures, were examined. Two different segmentation methods were used for validating the functionality of the proposed procedure. Results show that a statistically significant improvement in segmentation accuracy was achieved in each study case.

Comparing predictors of conversion to Alzheimer's disease using the disease state index.

Background: The Disease State Index (DSI) is a method which interprets data originating from multiple different sources, assisting the clinician in the diagnosis and follow-up of dementia diseases. Objective: We compared the differences in accuracy in differentiating stable mild cognitive impairment (S-MCI) and progressive MCI (P-MCI) obtained from different data combinations using the DSI. Methods: We investigated 212 cases with S-MCI and 165 cases with P-MCI from the Alzheimers Disease Neuroimaging Initiative cohort. Data from neuropsychological tests, cerebrospinal fluid, apolipoprotein E (APOE) genotype, magnetic resonance imaging (MRI) and positron emission tomography (PET) were included. Results: The combination of all parameters gave the highest accuracy (accuracy 0.70, sensitivity 0.71, specificity 0.68). In the different categories, neuropsychological tests (0.65, 0.65, 0.65) and hippocampal volumetry (0.66, 0.66, 0.66) achieved the highest accuracy. Conclusion: In addition to neuropsychological testing, MRI is recommended to be included for differentiating S-MCI from P-MCI. APOE genotype, CSF and PET may provide some additional information.

Using the Disease State Fingerprint Tool for Differential Diagnosis of Frontotemporal Dementia and Alzheimer's Disease

Background: Disease State Index (DSI) and its visualization, Disease State Fingerprint (DSF), form a computer-assisted clinical decision making tool that combines patient data and compares them with cases with known outcomes. Aims: To investigate the ability of the DSI to diagnose frontotemporal dementia (FTD) and Alzheimers disease (AD). Methods: The study cohort consisted of 38 patients with FTD, 57 with AD and 22 controls. Autopsy verification of FTD with TDP-43 positive pathology was available for 14 and AD pathology for 12 cases. We utilized data from neuropsychological tests, volumetric magnetic resonance imaging, single-photon emission tomography, cerebrospinal fluid biomarkers and the APOE genotype. The DSI classification results were calculated with a combination of leave-one-out cross-validation and bootstrapping. A DSF visualization of a FTD patient is presented as an example. Results: The DSI distinguishes controls from FTD (area under the receiver-operator curve, AUC = 0.99) and AD (AUC = 1.00) very well and achieves a good differential diagnosis between AD and FTD (AUC = 0.89). In subsamples of autopsy-confirmed cases (AUC = 0.97) and clinically diagnosed cases (AUC = 0.94), differential diagnosis of AD and FTD performs very well. Conclusions: DSI is a promising computer-assisted biomarker approach for aiding in the diagnostic process of dementing diseases. Here, DSI separates controls from dementia and differentiates between AD and FTD.