Integration of Neural Network Algorithms

Objective:

• Integrate artificial neural network segmentation into the NA-MIC kit for segmentation of upper extremity regions of interest.

Progress:

• Base algorithm has been developed using the annie neural network library
• Segmentation algorithm has been applied to segmentation of the phalanx bones on the index finger
  • Paper has been published on this initial work

To Do:

• Initial evaluation of this technique has been performed for phalanx bones.
• Need to rework ITK neural network libraries to allow the size of the network to be dynamic configured at run time instead of compile time
  • There are several issues that neeed to be resolved regarding the Neural network I/O in ITK
  • Consider the inclusion of the old ANN code that existed in the BRAINS software

Key Investigators:

• Iowa: Nicole DeVries, Nicole Grosland, Vincent Magnotta

Results

ANN Relative Overlap with Manual Rater

Subject	Proximal Phalanx	Middle Phalanx	Distal Phalanx
1	0.91	0.79	0.79
2	0.91	0.88	0.84
3	0.85	0.83	0.78
4	0.86	0.81	0.68
5	0.84	0.78	0.72
Average	0.87	0.82	0.76

Average Distance Between ANN Derived Surface and Physical Laser Scanned Surface

Subject	Proximal Phalanx (mm)	Middle Phalanx (mm)	Distal Phalanx (mm)
1	0.23	0.12	0.17
2	0.18	0.16	0.16
3	0.35	0.27	0.97
4	0.26	0.17	0.20
Bone Average	0.26	0.18	0.38

Links:

• Automated Bony Region Identification Using Artificial Neural Networks: Reliability and Validation Measurements. Esther E. Gassman, Stephanie M. Powell, Nicole A. Kallemeyn, Kiran H. Shivanna, Vincent A. Magnotta, Brian D. Adams, Nicole M. Grosland. Skeletal Radiology
• Musculoskeletal Imaging, Modeling and Experimentation (MIMX)

Figures:

Comparison of the manual and ANN segmentations of the phalanx bones. Manual segmentations are shown in red and automated in blue.