Heterogeneous Collection of Learning Systems for Confident Pattern Recognition

1. Heterogeneous Collection of Learning Systems for Confident Pattern Recognition Joshua R. New Knowledge Systems Laboratory Jacksonville State University

2. Outline • Motivation • Simplified Fuzzy ARTMAP (SFAM) • Interactive Learning Interface • System Demonstration • Conclusions and Future Work

3. Motivation

4. Motivation • Doctors and radiologists spend several hours daily analyzing patient images (ie. MRI scans of the brain) • The patterns being searched for in the image are standard and well-known to doctors • Why not have the doctor teach the computer to find these patterns in the images?

5. Motivation • Doctors and radiologists who use supervised AI systems for image segmentation: • Usually can not interactively refine the computer’s segmentation performance • Must be able to precisely select regions/pixels of the image to train the computer • Often do not use an interface that facilitates accomplishment of their task • Can easily lose where they are looking in the image when using magnification

6. Simplified Fuzzy ARTMAP (SFAM)

7. SFAM • In order to “teach the computer” to find tumors in neuro-images, a supervised machine learning system must be used • Simplified Fuzzy ARTMAP (SFAM) is a neural network that was created by Grossberg in 1987 and uses a mathematical model of the way the human brain learns and encodes information • This AI system was utilized because it allows very fast learning for interactive training (ie. seconds instead of days to weeks)

8. SFAM • SFAM is a computer-based system capable of online, incremental learning • Two “vectors” are sent to this system for learning: • Input feature vector gives the data is available from which to learn • Supervisory signal indicates whether that vector is an example or counterexample

9. SFAM • Data from which to learn • Feature vector from slice pixel values from shunted and single-opponency images (Whole Brain Atlas)

10. 0.35 0.90 Category 1 - 2 members Category 2 - 1 member y Category 4 - 3 members x SFAM • Vector-based graphic visualization of learning Array of Pixel Values

11. SFAM • Only one tunable parameter – vigilance • Vigilance can be set from 0 to 1 and corresponds to the generality by which things are classified (ie. vig=0.3=>human, vig=0.6=>male, 0.9=>Joshua New) 0.675 0.75 0.825

12. Category 1 - 2 members Category 2 - 1 member y Category 4 - 3 members x Vector 1 Vector 2 Vector 3 SFAM • SFAM is sensitive to the order of the inputs

13. SFAM • Voting scheme of 5 Heterogeneous SFAM networks to overcome vigilance and input order dependence • 3 networks: random input order, set vigilance • 2 networks: 3rd network order, vigilance ± 10%

14. SFAM

15. SFAM Threshold results Trans-slice results Overlay results

16. Interactive Learning Interface • Screenshot of Segmentation & Features

17. System Demonstration

18. Conclusions • Doctors and radiologists can teach the computer to recognize abnormal brain tissue • They can refine the learning systems results interactively • They can precisely select targets/non-targets • They can zoom for precision while maintaining context of the entire image • The interface developed facilitates task performance through display of segmentation results and interactive training

19. Future Work • Quantity of health-care can be increased by utilizing these trained “agents” to allow radiologists to only view the required images and directing their attention for the ones that are viewed • Quality of health care can be increased by using the agents to classify an entire database of images to highlight possibly overlooked or misdiagnosed cases