ADSA 02

Algorithm Development for Security Applications (ADSA) Workshop 2:

Workshop objectives

The second ADSA workshop, ADSA02, was held on October 7-8, 2009. Industry/practitioner, government and national lab participants were: Optosecurity, Reveal Imaging, Telesecurity Sciences, L-3 Communications, Optosecurity, Surescan, Analogic, GE Security, Mercury Computers, Guardian Technologies, Siemens Corporate Research. Department of Homeland Security, Lawrence Livermore National Laboratory, Massachusetts General Hospital, Transportation Safety Administration and Pacific Northwest National Laboratory.

This workshop was conducted to discuss the implementation of a challenge for segmenting objects of interest (OOI) from volumetric CT scans of baggage. Segmentation and classification are the two steps that are usually found in algorithms that perform automated threat recognition (ATR). Only the segmentation step of ATR is of interest for this grand challenge.

The objectives of the workshop were to discuss the following aspects of executing the grand challenge:

• CT segmentation grand challenge definition
• Dataset creation
• Participant identification
• Entry criteria and funds allocation
• Segmentation algorithm development and testing
• Independent validation and testing of the segmentation algorithms
• Demonstration of algorithms
• Creation of final report

Workshop outcomes

The main outcomes of the workshop are as follows:

• The grand challenge for segmenting OOIs from volumetric CT images should be performed.
• A number of refinements to the grand challenge were suggested.
• There are relevant precedents in the medical imaging and other communities(e.g., the Netflix grand challenge) that should be researched in order to follow their best practices.

• It is speculated that advanced reconstruction algorithms will have a bigger impact on the performance of CT-based explosives detection equipment compared to advances in segmentation. However, a prerequisite for developing reconstruction algorithms is having segmentation algorithms available in order to assess the impact of improved image quality on segmentation. Therefore, it may be necessary to complete the grand challenge for CT segmentation before implementing a grand challenge for reconstruction.
• Third parties have begun to work on the problems described in the two workshops to date. This is due, in part, to disclosing problem statements and potentially making databases available via NDA access.

A copy of this report can be requested by contacting Mariah Nobrega at 617-373-3031 or via email mnobrega@coe.neu.edu