Monday March 30 Sancy and Steve decided that Tracy would devote one specific day per week to his Derm activities, rather than Tracy working in bursts to meet particular Derm deadlines, like grants or phone conferences. Tracy selected Mondays as his Derm Day.
This Monday we (Tracy and Steve) outlined the project to bring high-performance computing to evaluation of Dermoscopy images.
Here is the project in brief:
1. A dermatologist sends a dermoscopy image to our website
2. The website sends the image to our computer with GPU bank for high-speed analysis.
3. The computer return the analysis to the website which returns it to the dermatologist.
ANALYSIS: The analysis interprets the color balance of the image pixels in terms of blood and melanin at various depths in the lesion, and variation in optical scattering properties (i.e., epidermal proliferation, dermal fibrosis).
We will approach INTEL about installing a bank of their xenon-pi GPUs on a website to launch the project. We will pursue NIH funding as a resource center to support the website.
As a pilot project, we are analyzing this melanoma image (see figure 1) from Debbie. The strategy is to use spatial frequency filtering to identify the low-frequency patterns due to deeper tissue structures (blood, melanin) versus high-frequency patterns due to superficially located structures. The color balance (Blue/Red, Green/Red) will indicate the depth of the pigments. This strategy is only a first approach. There are more sophisticated approaches that will be pursued as well.
FIGURE 1: (LEFT) Dermoscopy image. (RIGHT) Red, Green and Blue channels of image, filtered for low, medium, and high spatial frequency content.
Vessy will slice the lesion and create a set of histopathology slides, to be photographed. These images will guide the creation of a 3D computer simulation of the lesion, which will generate the equivalent dermoscopy color image.
The above analysis will also be extended to the images acquired by Dan Gareau's Spectral Camera. Steve discussed with Dan about how to anonymously share the images generated by the camera. His camera automatically uploads the images to Rockerfeller Univ., and we need to gain access to the images.
We have our own Spectral Camera (specCAM, built by Ravi Samatham) which can operate at 30 wavelengths between 400-700 nm (purple to deep red). We will use the specCAM to develop the dermoscopy analysis.