Albert Montillo PhD, University of Pennsylvania   Research Associate Computational Biomedicine, Imaging and Modeling Center Phone: (267) 257-5094

Home Research Publications Lectures Professional Service Short Bio

 Links to our research areas:

• Large data analytics and machine learning
• Medical image analysis
• Computer vision and image processing
• Fast marching methods for implicit modeling

• Large data analytics and machine learning

  Feature selection and dimensionality reduction

  We are interested in optimal pairing of feature selection methods with state-of-the-art regression machines in ways that maximize performance and are insensitive to their parameterization. We study four effects of pairing including: learning time, model selection time, prediction time, and prediction accuracy. Five feature selection methods are tested, including one that we propose based on a combination of forward selection and correlation relevancy metrics. The proposed method yields hundreds of times performance improvement for both support vector regression and Random Forest regression over a wide parameter ranges.

  

   

  Discriminative models for image-based regression

  We are interested in highly accurate discriminative models to drive image-based regression and provide near instantaneous prediction on novel test images. In one project, we propose a method to automatically predict the age of a person given only a single image of a person's face. We propose new feature construction methods, including a new, efficient, region descriptor called the Gabor integral histogram and use the features to employ the Random Forest regressor to learn a discriminative model of aging. We tested our method for accuracy on publicly available face data, and found a prediction accuracy of 3.1 years, which compares favorably to human prediction on the data. The method also provides a model which is readily interpretable, includes importance measures for each feature, and finds features pertinent for facial development without using a prior anthropometric model.

.

• Medical image analysis

Biomechanical models for tissue segmentation and tracking

Most of the clinically relevant parameters that characterize heart function can be readily derived from a dynamic segmentation of the heart in 3D+time images. Obtaining the dynamic segmentation is challenging due to a wide variety of heart shape, complex motion, and relatively low data quality. We propose a method that uses a volumetric deformable model with regularization constraints derived from biomechanics to dynamically segment the left and right ventricles of the human heart. While most other methods start from a manual segmentation, our method automatically constructs the patient specific model using adaptive statistics-driven, image processing and shape based interpolation. By progressively incorporating elastic tissue properties the method then achieves high segmentation accuracy during the heartbeat of a few mm for both normal and diseased hearts.

Tissue tracking using a Gabor filter bank

We are interested in tracking complex tissue motion. We propose a new approach for the extraction of tissue deformation based on the responses of a bank of Gabor filters. Additionally we propose an interpolation method to recover all information at a finer resolution than the filter bank parameters. The method is fully automated; requiring no user supervision. Tests of our method on synthetic images of tissue undergoing an isovolumic contraction, for which we have exact ground truth, showed very rapid computation (a few seconds) and sub-pixel tracking accuracy.

(a-b) Isovolumic contracting short axis (synthetic) image with (c) automatically recovered displacement and (d) resulting small displacement error

Generative models for segmentation and registration

Spatially parameterized segmentation of the whole brain

We develop a supervised learning algorithm to build a 3D statistical brain model in MRI. Recognizing that the imaged intensities of each structure vary spatially even within the same tissue, we construct a learning method that uses spatially parameterized Bayesian classifiers with a Markov random field label model for label relaxation. Compared to most brain segmentation methods which label only 3 or 4 structures such as CSF, white matter, gray matter and bone, our method accurately label over 35 structures throughout the brain. We tested our method against a team of expert anatomists and show how the volumes of 15 subcortical can be used to detect morphological changes that presage Alzheimer's disease.

 

Sample 3D brain labeling result, and detail showing labeling of subcortical region:

 

Building a probabilistic a model of 3D registration

Accurate registration of complex 3D structures often involves a very high dimensional nonlinear transformation requiring hundreds of thousands of parameters. Elastic matching regularizes the registration using elasticity constraints typically combining the finite element method and the theory of continuum mechanics. Recognizing that most deformable shapes change along a low number of deformation modes, we use elastic matching to train a probabilistic deformation mode, and investigate run-time performance gains by deforming first along eigen-deformation modes.

---

• Computer vision and image processing

Recovering the geometric description of an object from an image

We propose a method to rapidly and accurately estimate the orientation of an N-dimensional object by processing the patterns in its spatial frequency transform using a non-linear scaling and Hough-line finding. We then refine a geometric description of an object using a hierarchical estimation of object feature dimensions and projected intensity profile processing. US patents #6813377 and #6526165 have been issued for these methods, which are now deployed in successful industrial inspection applications.

 

Segmenting regions of similar texture

Some classes of objects have at least a portion of which has a characteristic pattern or texture. In order to rapidly recover the pos of such objects, we propose a method that employs spatial frequency analysis to construct a feature vector for every pixel, including: frequency spectrum, power spectrum, and angles of dominant powers. We integrate this information to segment regions of similar texture. We validate the method on a variety of industrial inspection applications. US patent #6647132 has been issued for this method.

 

Adaptive, non-stationary image denoising

Clinical imaging protocols are highly variable between imaging sites. In MR imaging, images are affected by additional variations include the number and placement of surface coils. To normalize image variation, we create a non-stationary noise suppression method that adaptively regulates anisotropic diffusion using the strength of an estimated intensity inhomogeneity field. We test our method on synthetic images, simulating surface coils using Biot-Savart electromagnetics law and on thousands of clinical MR images. Statistical tests reveal that the edge-strength of critical tissue boundaries are significantly more well preserved using our adaptive denoising.

 

 

Optimal correction of intensity inhomogeneity

The intensity of a tissue often varies spatially across an image. Correction of this inhomogeneity is an important step towards automating image analysis. Such inhomogeneity is particularly problematic in MRI, where higher magnetic fields increase image resolution but tend to come with increased intensity inhomogeneity. In 3D the inhomogeneity usually varies in all three dimensions. In 3D+time images, the intensity can also vary over time as magnetized blood moves into or out of the imaged region. We propose iterative methods to progressively estimate large single tissue regions, estimate the inhomogeneity and suppress the inhomogeneity on 3D+time images. Additionally we investigate the effect of the order of denoising and inhomogeneity correction. On tests of over 1000 MR images we find that for the greatest reduction in intensity variation, inhomogeneity correction should precede noise suppression.

.

---

• Fast marching methods for implicit modeling

Optimal time Eikonal solution on artifact-laden triangulated manifolds

Solving the Eikonal equation can facilitate many types of image analyses and simulations. Most solutions require special processing to handle artifacts such as holes and obtuse triangles. We present an optimal time solution for triangulated surfaces that is an extension of the Fast Marching method. Our method handles all mesh artifacts uniformly, provides exceptionally low parameter sensitivity. Tests on a wide range of examples from computer graphics and medical imaging, demonstrate the high accuracy and suitability for geodesic measurements and mesh construction.

Implicit 3D curvature as a new speed term for level sets

We explore the use of local shape properties to constrain level set shape segmentation. Certain anatomic shapes, such as vasculature, contain consistent local shape properties. In particular these shapes have high first principal curvature while low second principal curvature, while the ratio of these curvatures tends to be relatively constant. We compute the principle curvatures, k1 and k2, directly from the gaussian curvature, K, and mean curvature, H, which in turn are computed from 3D image derivatives.

Comparison of shape representation methods

We are interested in methods that efficiently represent complex 3D shapes. We study and compare four competing methods including: (1) two coupled surface propagation methods with applications to brain segmentation, (2) one method for geodesic active contours constrained by a generative statistical shape model with applications to object segmentation, and (3) a medial axis shape representation method. The advantages and disadvantages are detailed and promising future directions are discussed.