Current Projects

Identification of Autism from fMRI

Classification Network

Prediction of Autism Treatment Outcome

Treatment Outcome Prediction

Body Motion Correction in Dynamic PET Imaging

Past Projects

Classification of Breathers for PET Imaging

Respiratory Traces

Positron-emission tomography (PET) imaging plays an important role in cancer management. Motion correction is essential for mitigating respiratory motion for accurate tumor detection and quantification in the lung and abdomen. Existing motion correction methods based on phase gating work well for “regular” breathers, but result in blurry images for “irregular” breathers due to inter-cycle and intra-cycle motion variations. Sophisticated motion correction techniques work well for all breathers, but their high computational complexity dictate that such approaches be reserved for irregular breathers. To handle this motion correction problem, we propose to first classify regular and irregular breathing patterns using respiratory trace data acquired during the PET scan. The respiratory traces are used as input to a deep learning model based on recurrent neural networks to learn the breather type directly from the trace signal. The breather classification system will provide the foundation for a personalized motion correction framework to provide the optimal tumor imaging for each individual patient.

Fast Single Particle Reconstruction for Cryo-Electron Microscopy

Cryo-electron Microscopy (cryo-EM) has emerged as a key tool for studying protein structure. The 2D cryo-EM projection images of frozen protein samples, i.e. particles, are used to infer the protein’s 3D structure through a process called single particle reconstruction. Single particle reconstruction methods based on the maximum-likelihood principle are popular because of their ability to produce high resolution structures. However, these algorithms are computationally very expensive, requiring a network of servers. To address this problem, we developed a new algorithm called SubspaceEM for accelerating maximum-likelihood reconstructions by utilizing subspace approximations of the cryo-EM images and the structure projections. The speedup is by orders of magnitude, and the new algorithm produces similar quality reconstructions compared to the traditional maximum-likelihood formulation.

Image Registration with Missing Correspondences

Classification Network

Medical image registration is used to align patient images from different treatment time points to a single reference frame. This process of establishing correspondences between images is important for making meaningful comparisons across scans. Traditional registration algorithms assume a one-to-one correspondence between features in the images to be aligned. However, this assumption is clearly violated when the images are missing correspondences, which often occurs when dealing with patient data due to treatment effects or disease progression. Standard registration methods, therefore, will likely fail to align actual corresponding features, especially near locations with missing data, i.e. the regions of interest. My PhD dissertation focused on the development of an automated image registration algorithm to deal with the missing correspondence problem. The key idea is to incorporate the estimation of a label map segmenting the valid and missing correspondence voxels during the registration. We pose the registration problem in a maximum a posteriori framework and jointly solve for the transformation parameters and label map. The developed algorithmic framework is general and can be adapted to many missing correspondence problems, including alignment of preoperative and postresection brain images and tracking brain metastases over time.