Project Goal
The overall goal is to speed up protein crystallization efforts, and hence discover new 3D structures. A single robot can set up and photograph about 100,000 crystallization experiments per day, but currently interpretation of the images is mostly manual.
Figure 1: Precipitate image.
Figure 2: Microcrystal image.
Figure 3: Mixture image.
Manual interpretation of individual images is quite good-- much better than computer vision can offer-- but manual interpretation of a suite of 100's or 1000's of images with 3 or 4 varying chemical parameters is not very good.
Project Scope
The most important subproblem is distinguishing precipitates from microcrystals. This problem is hard even for a person (see the attached images: one precipitate, one microcrystal, and one mixture). The project involves interfacing with Dr. Marshall Bern of PARC, and indirectly with some scientists at the Scripps Research Institute and the Joint Center for Structural Genomics.
Tasks
The project will be accomplished through the following tasks.
- Task 1: Assemble some training and test sets out of the more than 1 million images at JCSG.
- Task 2: Develop an algorithm for segmenting the drop boundary. We currently have such an algorithm at PARC (about 93% accurate), and that is good enough for now. But we suspect that smarter computer vision methods (e.g., level-set methods) might be more accurate, so the student may choose to spend some time on this task as well as the more crucial Task 3.
- Task 3: Develop statistical features for the interiors of drops that correlate well with human notions such as "grain size" and "number of crystals". Possible tools include wavelets and over-complete wavelets.
- Task 4: Develop a precipitate versus microcrystal binary classifier.
Project Status
Samuel Cheng, Shaohua Sun, and Xiaoqing ZhuPoint of Contact
Daniel RussakoffMidterm Report
submittedFinal Report
submitted
