Research

I haven’t done academic research in a while, so I thought I’d summarize it somewhere. Here’s as good a place as any. For my more up-to-date interests and work, please see my resume.

1. •2001-2002: I worked at Rockefeller Univerity’s Bio-Imaging Resource Center where I etched my name into a cell, and learned a lot about Microscopy.
   

2. •2002-2003: I moved to Rockefeller’s GENSAT project, where I was a sysadmin and I developed some semiautomated image processing software called ImagePrep.
   

3. •2003: I worked at NYU Medical Center on some preliminary research to see if we could use an automated technique based on texture analysis to detect small airway disease from low-dose lung CT scans. I presented my paper, titled "Automated assessment of small airway disease from low-dose lung CT: a preliminary study", at SPIE's 2003 Physics of Medical Imaging conference.
   

4. •2005: I worked at Emory School of Medicine’s Department of Rehabilitation Medicine on a project involving PET scans of brains while subjects were involved in tasks requiring them to look through mirrors.

Black and White

Low dose lung CT of a diseased lung

Color Separation

Research Related Photos (click to enlarge)

Photobleaching

The picture above was taken with a Zeiss LSM 510 Confocal Microscope, The fluorescent chemicals used to label parts of the cell were bleached out by a high powered laser that scanned across the cell, in much the same way that an electron beam scans across a TV screen. After photo-bleaching, the laser was switched to a lower power mode to acquire the image. For scale, the width of a human hair may range from 17µm to 180µm, depending on who you ask. I find it interesting that the dots of my umlaut in this cell are much smaller than the width of my hair, at roughly 10µm.

Full Photo

ImagePrep, the software I wrote while working at GENSAT was designed to speed up image preparation, which was previously done in Photoshop. Since the images were very large, simple operations as image rotation could take up to one and a half minutes using Photoshop. ImagePrep allowed rotation to be performed on thumbnails, so that rotation of the full image could be done unattended, along with automatic contrast and brightness settings and automatic removal of extraneous data in the images, such as dirt and debris. All imageprep computation was performed without intermediate storage resulting is less computational error, and distortion. (This is a commonly overlooked problem with Photoshop and other image manipulation software.)

The image cleanup algorithm was based on an algorithm by Binsheng Zhao et. al. 1. Their algorithm, originally designed for isolating small pulmonary nodules in 3-D lung CT images, was modified for 2D bright-field microscope images. It has also been modified to improve speed. My modified algorithm is a simple 2-step process that first determines an optimal threshold using an accelerated gradient-based technique, and then produces a mask from the threshold. The mask is then cleaned by elementary binary image operations. The process is conservative but effective at removing most debris.

1 Binsheng Zhao, David Yankelevitz, Anthony Reeves and Claudia Henschke, "Two-dimensional multi-criterion segmentation of pulmonary nodules on helical CT images," Med. Phys. 26 (6). 889-895 (1999).