Machine Learning for Brain/Cardiac MRI Analysis
Brain MRI Registration with Learned Anatomical Variation (Medical Image Analysis 2010, MICCAI 2012, MedIA-MICCAI Best Paper Award)
Fig. 1. Left: geodesic path of cortical surface from the moving to the template. Note the relatively gradual change of cortical patterns through the paths. Right: comparison of the final registration results from the geodesic versus the direct registration. Our results present more smooth and realistic warping of the cortices when compared to the unnatural warping from the direct method. Representative regions in which the geodesic registration is markedly better are shown by cyan-colored circles.
Medical image registration is the process of aligning images of multiple subjects to compensate for inter-subject differences and to establish correspondence between every voxel of the images. Accurate registration is a challenging problem, especially when there is large variation in the anatomies (e.g., cerebral cortex). To tackle this challenge, we proposed a novel framework for large deformation registration using the learned manifold of anatomical variation in the data. Specifically, we developed the technique which decomposes a large deformation between two images into a series of small deformations along the geodesic path on the manifold of anatomical variation. Using this technique, I was able to visualize the major anatomical variation of the parcellated cortical surface and achieved significant improvement in registration accuracy over conventional direct registration as shown in Fig. 1.
Disease Classification in Brain/Cardiac MRIs (PRNI 2011, IEEE TMI 2014, PRNI Student Travel Award)
Fig. 2. Manifold embeddings of brain MRIs with Alzheimer' Disease (AD), converted Mild Cognitive Impairment (cMCI), non-converted Mild Cognitive Impairment (ncMCI), and Control Normal (CN). The embeddings conveniently summarize the change in ventricle size and gray matter density as the most dominant parameters, while separating AD/cMCI patients from CN/ncMCI subjects.
Manifold learning generates low-dimensional representations that nonlinearly embed high-dimensional images. The resulting low-dimensional representations can encode biologically relevant variation in images, and can be used to facilitate disease classification. For example, we feed embeddings of brain MRIs into a semi-supervised classifier to predict conversion from MCI to AD. As illustrated in Fig. 2, the embeddings can encode biological variation such as ventricle size and gray matter density, while locating cMCI/ncMCI nearby AD/CN, respectively. Therefore, embeddings in conjunction with the semi-supervised classifier can boost the early detection performance of MCI to AD conversion.
Fig. 3. Examples of classified cardiac MRI scans with reconstructive surgery of Tetralogy of Fallot (TOF). This figure illustrates the progression of the shape of the right ventricle from being round in normal controls to the different degrees of dilation and irregularity in TOF patients. This indicates that our classification can capture the shape abnormalities in the heart.
In identifying the impact of a disease from cardiac MRIs, cardiologists often rely on volumetric measurements of the specific anatomical regions (e.g., left/right ventricular volume of the heart). However, such volumetric measurements have not shown high sensitivity and specificity in diagnosis of individuals because the spatial pattern of pathology is complex. Therefore, we used manifold learning to generate low-dimensional shape encodings from deformation-based descriptors. By feeding the manifold-based shape encodings into a supervised classifier, I was able to detect shape changes induced by reconstructive surgery for the congenital heart defect Tetralogy of Fallot (TOF) as shown in Fig. 3.
Brain Tumor Segmentation using Multimodal Image Synthesis (MICCAI 2013, IEEE TMI 2015, Multimodal Brain Tumor Segmentation Challenge Winner)
Fig. 4. Brain Tumor Detection using synthesis of "pseudo-healthy" T2 image from T1. The abnormality maps are computed by subtracting the synthesized T2 from the original T2. Our abnormality maps highlight tumor core and edema while subduing areas not related to tumor (e.g., ventricle).
MRI provides a multitude of image modalities, each locally quantifying different characteristics of the underlying anatomy. For the particular brain tumor, pathological tissue enhances in T2 yielding a hyper-intense appearance while it does not substantially alter the intensity profile in (non-contrasted) T1. Then, we build a system for abnormality detection in which "pseudo-healthy" T2 image is synthesized from input T1 image via patch-based label propagation employing a database of T1/T2 pairs of healthy subjects. The abnormality map is computed by subtracting the synthesized T2 from the original T2 image, showing prominent tumor regions as shown in Fig. 4. Our approach provides a good initial guess for further tumor segmentation steps such as tissue classification random forests.