Automatic Target Recognition (ATR) for CT-based Airport Screening System

Description: Project to classify CT images as target or non-target.

EECE 5890: Machine Learning for Image Processing

Professor: Dong Hye Ye, Ph.D.

Authors

• John Hattas
• Nathan Lang

Resources

Python

NiBabel - read / write access to some common neuroimaging file formats
PyNifti - older version of NiBabel

Visualization

MIPAV - quantitative analysis and visualization of medical images

Details

Targets

• Saline
  • 3.5%, 10%, 15% concentrations
• Modeling (polymer) clay
• Rubber sheets: ¼” thickness (minimum) + other rubber in bags

Non-Targets

• Food
• Drinks
• Electronics
• Magazines
• Containers not filled with saline

Dataset

Cropped CT image for each segmented object

Target labels

Non-target:0, Saline:1, Rubber:2, Clay:3

Image Quality

Many artifacts which lead to imprecise density, volume, mass, shape

Possible features

• Mass
• Mean
• Standard deviation
• Histograms
• Higher-order moments
  • Skew, kurtosis, entropy
• Texture
  • Wavelets

Possible classifiers

• PCA
• SVM
• Decision Tree
• Adaboost
• Deep neural network

Performance metric

PD = # targets detected / # targets scanned
PFA = # false alarm objects / # non-targets scanned

Goal: PD > 90%, PFA < 10%

Data Files

Nifti file format: Standard Neuroimaging File Format
.nii.gz: gzipped image