Mikko Lilja

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.

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.

Vessel segmentation for ablation treatment planning and simulation

In this paper, a novel segmentation method for liver vasculature is presented, intended for numerical simulation of radio frequency ablation (RFA). The developed method is a semiautomatic hybrid based on multi-scale vessel enhancement combined with ridge-oriented region growing and skeleton-based postprocessing. In addition, an interactive tool for segmentation refinement was developed. Four instances of three-phase contrast enhanced computed tomography (CT) images of porcine liver were used in the evaluation. The results showed improved accuracy over common approaches and illustrated the methods suitability for simulation purposes.

COMPARATIVE EVALUATION OF VOXEL SIMILARITY MEASURES FOR AFFINE REGISTRATION OF DIFFUSION TENSOR MR IMAGES

Deriving an accurate cost function for tensor valued data has been one of the main difficulties in diffusion tensor image (DTI) registration. In this work, we evaluate and compare five voxel similarity measures: Euclidean distance (ED), Log-Euclidean distance (LOG), distance based on diffusion profiles (DP), diffusion mode based similarity (MBS), and multichannel version of sum of squared differences (SSD). In evaluation we used an optimization-independent evaluation protocol to assess the capture range, the number of local minima, and cyclic registrations to evaluate consistency. Statistically significant differences were observed: DP and MBS were found to be the most consistent similarity measures, ED had the least number of local minima, and SSD was inferior to other similarity measures in all evaluations.

Automatic extraction of mandibular bone geometry for anatomy-based synthetization of radiographs

We present an automatic method for segmenting Cone-Beam Computerized Tomography (CBCT) volumes and synthetizing orthopantomographic, anatomically aligned views of the mandibular bone. The model-based segmentation method was developed having the characteristics of dental CBCT, severe metal artefacts, relatively high noise and high variability of the mandubular bone shape, in mind. First, we applied the segmentation method to delineate the bone. Second, we aligned a model resembling the geometry of orthopantomographic imaging according to the segmented surface. Third, we estimated the tooth orientations based on the local shape of the segmented surface. These results were used in determining the geometry of the synthetized radiograph. Segmentation was done with excellent results: with 14 samples we reached 0.57 ± 0.16 mm mean distance from hand drawn reference. The estimation of tooth orientations was accurate with error of 0.65 ± 8.0 degrees. An example of these results used in synthetizing panoramic radiographs is presented.

Fast and accurate voxel projection technique in free-form cone-beam geometry with application to algebraic reconstruction

A modification is introduced to an established ray-tracing technique for computing line integrals in a rectangular voxel lattice in a free-form cone-beam imaging geometry. The result is a fast and accurate technique for computing a digitally reconstructed radiograph (DRR), which can easily be extended to simultaneous algebraic reconstruction technique (SART). The proposed method is described in detail and simulated experiments show that high resolution DRR images and algebraic reconstructions can be computed in a very reasonable time even without parallel computation. The proposed technique is highly feasible for further development from the perspective of clinical imaging applications.

Model-Based Segmentation of Reconstructed Dental X-Ray Volumes

Modern reconstruction algorithms allow volumetric imaging with conventional 2D dental X-ray systems. Volumetric images are useful in dental implantology, where the correct identification of key structures such as the edges of the mandible and the mandibular nerve is critical. This paper presents a segmentation method capable of extracting the mandible. The segmentation is based on a statistical model which was first transformed affinely and finally deformed non-rigidly to the object. The method was tested on three volumes with good results: mean distances between the deformed and manually segmented reference surfaces were 0.26, 0.34 and 0.50 mm. Applications of the method include the extraction of slices orthogonal to the mandibular bone centerline and local, anatomy based image enhancement.

Brain extraction from MR images using a combination of segmentation fusion and marker-controlled watershed transform

A new brain extraction method for MR images is presented that combines segmentation fusion with an active segmentation step. Three common brain extraction methods (Brain Extraction Tool (BET), 3DSkullStrip and FreeSurfer) were used to provide the input segmentations. The areas where the input segmentations agreed were fused normally, while the areas of disagreement were left to be handled by an active segmentation in a marker-controlled watershed framework. The performance of the proposed algorithm was compared with the input segmentations as well as with the majority voting and staple meta-algorithms. Three evaluation criteria related to the overlap error, average distance and volume differences were used on two datasets. The results showed that the proposed method outperformed the input segmentations as well as the meta-algorithms on all the evaluation criteria for both datasets. It is concluded that the proposed method appears highly suitable for large studies involving brain extraction, where full automation and robust segmentation results are needed for all datasets.

An enhanced version of ITK-SNAP for preoperative inspection and refinement of surface mesh models

This paper describes a modified and enhanced version of the GPL-licensed, ITK-/VTK-based segmentation software ITK-SNAP. The modified software is entitled IMPPACT-SNAP and was motivated by the requirement for preoperative inspection and manual refinement of surface mesh liver segmentations in the research project IMPPACT. In this context, the liver segmentations are used for simulating the radiofrequency ablation of liver tumors to optimize the outcome of the treatment. The implemented modifications add the functionality of visualizing and manually deforming surface mesh segmentations and improve the navigational features of the software. The result is an easy-to-use tool that builds on ITK-SNAPs user-oriented GUI and framework.