An Advanced User Interface for Pattern Recognition in Medical Imagery: Interactive Learning, Contextual Zooming, and Gesture Recognition Joshua R. New Knowledge Systems Laboratory Jacksonville State University Knowledge Systems Lab JN 5/29/2016 Outline • Introduction • Techniques: Segmentation, Magnification, Exploration • Solutions: – Interactive Learning – Contextual Zooming – Gesture Recognition • Conclusions Knowledge Systems Lab JN 5/29/2016 Introduction Medical imagery… • Consists of millions of images produced annually which doctors must gather and analyze • Entails several modalities for each patient, such as MRI, CT, and PET Refine techniques for facilitating comprehension of this data Knowledge Systems Lab JN 5/29/2016 Outline • Introduction • Techniques: Segmentation, Magnification, Exploration • Solutions: – Interactive Learning – Contextual Zooming – Gesture Recognition • Conclusions Knowledge Systems Lab JN 5/29/2016 Techniques • Common techniques for facilitating data comprehension: – Segmentation – Labeling of images – Magnification – Precision viewing – Exploration – Interacting intuitively with complex, 3D data Knowledge Systems Lab JN 5/29/2016 Why Segmentation? • Doctors and radiologists: – Spend several hours daily analyzing patient images (ie. MRI scans of the brain) – Search for patterns in images that are standard and well-known to doctors • Why not have the doctor teach the computer to find these patterns in the images? Knowledge Systems Lab JN 5/29/2016 Why Magnification? • Doctors and radiologists: – Must be able to precisely view and select regions/pixels of the image to train the computer – Can easily lose where they are looking in the image when using magnification • Why not use visualization techniques to preserve context while allowing precise selections? Knowledge Systems Lab JN 5/29/2016 Why Exploration? • Doctors and radiologists: – Need to intuitively interact with the system to maximize task performance – Need to perform this interaction while being unencumbered • Why not use vision-based recognition to allow interaction with the data? Knowledge Systems Lab JN 5/29/2016 Outline • Introduction • Techniques: Segmentation, Magnification, Exploration • Solutions: – Interactive Learning – Contextual Zooming – Gesture Recognition • Conclusions Knowledge Systems Lab JN 5/29/2016 Problems & Solutions • Problem #1: Segmentation • Solution #1: Interactive Learning • Problem #2: Magnification • Solution #2: Contextual Zoom • Problem #3: Exploration • Solution #3: Gesture Recognition Knowledge Systems Lab JN 5/29/2016 Platform • Med-LIFE: – “L”earning of MRI image patterns – “I”mage “F”usion of multiple MRI images – “E”xploration of the fusion and learning results in an intuitive 3D environment • Images used from “The Whole Brain Atlas” – http://www.med.harvard.edu/AANLIB/home.html Knowledge Systems Lab JN 5/29/2016 Outline • Introduction • Techniques: Segmentation, Magnification, Exploration • Solutions: – Interactive Learning – Contextual Zooming – Gesture Recognition • Conclusions Knowledge Systems Lab JN 5/29/2016 Simplified Fuzzy ARTMAP • Simplified Fuzzy ARTMAP (SFAM) – An AI neural network (NN) system – Capable of online, incremental learning – Takes seconds for tasks that take backpropagation NNs days or weeks to perform Knowledge Systems Lab JN 5/29/2016 Vector-based Learning • Two “vectors” are sent to this system for learning: – Input feature vector provides the data from which SFAM can learn – ‘Teacher’ signal indicates whether that vector is an example or counterexample Knowledge Systems Lab JN 5/29/2016 Feature Vector • Pixel values from images (16 for each slice) Knowledge Systems Lab JN 5/29/2016 Learning Visualization • Vector-based graphic visualization of learning Category 1 - 2 members y Array of Pixel Values 0.30 0.45 Category 2 - 1 member Category 4 - 3 members x Knowledge Systems Lab JN 5/29/2016 Learning Associations Full Results T2Results Detailed Knowledge Systems Lab JN 5/29/2016 Varying Vigilance • 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 Knowledge Systems Lab JN 5/29/2016 Input Order Dependence • SFAM is sensitive to the order of the inputs Vector 1 Category 1 - 2 members Vector 2 Category 2 - 1 member y Vector 3 Category 4 - 3 members x JN 5/29/2016 Knowledge Systems Lab Heterogeneous Network • 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% Knowledge Systems Lab JN 5/29/2016 Network Segmentation Results Knowledge Systems Lab JN 5/29/2016 Segmentation Results Threshold results Trans-slice results Overlay results Knowledge Systems Lab JN 5/29/2016 Segmentation Screenshot Knowledge Systems Lab JN 5/29/2016 Interactive Learning System Demonstration Knowledge Systems Lab JN 5/29/2016 Segmentation Solution • Doctors and radiologists: – Spend several hours daily analyzing patient images (ie. MRI scans of the brain) – Search for patterns in images that are standard and well-known to doctors • Solution: – Doctors and radiologists can teach the computer to recognize abnormal brain tissue – They can refine the learning systems results interactively Knowledge Systems Lab JN 5/29/2016 Outline • Introduction • Techniques: Segmentation, Magnification, Exploration • Solutions: – Interactive Learning – Contextual Zooming – Gesture Recognition • Conclusions Knowledge Systems Lab JN 5/29/2016 Zooming Approaches Inset Overlay Chip Window Knowledge Systems Lab JN 5/29/2016 Research & Business • Carpendale PhD Thesis – Elastic Presentation Space – rubber sheet images via mathematical constructs • IDELIX (www.idelix.com) – Pliable Display Technology – software development kit (SDK) product – Boeing: 20% increase in productivity Knowledge Systems Lab JN 5/29/2016 Zoom Visualization Contextual Zoom Wireframe View Knowledge Systems Lab JN 5/29/2016 Contextual Zoom System Demonstration Knowledge Systems Lab JN 5/29/2016 System Comparison Previous System Zoom Overlay Contextual Zoom Knowledge Systems Lab JN 5/29/2016 Magnification Solution • Doctors and radiologists: – Must be able to precisely view and select regions/pixels of the image to train the computer – Can easily lose where they are looking in the image when using magnification • Solution – They can precisely select targets/non-targets – They can zoom for precision while maintaining context of the entire image – The interface facilitates task performance through interactive display of segmentation results Knowledge Systems Lab JN 5/29/2016 Outline • Introduction • Techniques: Segmentation, Magnification, Exploration • Solutions: – Interactive Learning – Contextual Zooming – Gesture Recognition • Conclusions Knowledge Systems Lab JN 5/29/2016 Motivation • Gesturing is a natural form of communication: – Gesture naturally while talking – Babies gesture before they can talk • Interaction problems with the mouse: – Have to locate cursor – Hard for some to control (Parkinsons or people on a train) – Limited forms of input from the mouse Knowledge Systems Lab JN 5/29/2016 Motivation • Problems with the Virtual Reality Glove as a gesture recognition device: – Reliability – Always connected – Encumbrance Knowledge Systems Lab JN 5/29/2016 System Diagram User Rendering Hand Movement Update Object User Interface Display Gesture Recognition System Image Capture Image Input Standard Web Camera Knowledge Systems Lab JN 5/29/2016 System Performance • System: • OpenCV and IPL libraries (from Intel) • Input: • • 640x480 video image Hand calibration measure • Output: • • • • Rough estimate of centroid Refined estimate of centroid Number of fingers being held up Manipulation of 3D skull in QT interface in response to gesturing Knowledge Systems Lab JN 5/29/2016 Calibration Measure • Max hand size in x and y orientation (number of pixels in 640x480 image) Knowledge Systems Lab JN 5/29/2016 Saturation Extraction Saturation Channel Extraction (HSL space): Original Image Hue Lightness Saturation Knowledge Systems Lab JN 5/29/2016 Gesture Recognition Pipeline Knowledge Systems Lab JN 5/29/2016 Gesture Recognition Pipeline Knowledge Systems Lab JN 5/29/2016 Gesture Recognition Pipeline a) 0th moment of an image: M 00 I ( x, y) b) 1st moment for x and y of an image, respectively: M 10 x I ( x, y) M 01 y I ( x, y) c) 2nd moment for x and y of an image, respectively: 2 M 20 x 2 I ( x, y) M 02 y I ( x, y) d) Orientation of image major axis: arctan M 2 1 1 xc yc M 00 M 2 0 x 2 M 0 2 y 2 c c M M 00 00 2 Knowledge Systems Lab JN 5/29/2016 Gesture Recognition Pipeline Radius 0.19 * ( HandSizeX HandSizeY ) • The finger-finding function sweeps out a circle around the rCoM, counting the number of white and black pixels as it progresses • A finger is defined to be any 10+ white pixels separated by 17+ black pixels (salt/pepper tolerance) • Total fingers is number of fingers minus 1 for the hand itself Knowledge Systems Lab JN 5/29/2016 System Setup System Configuration System GUI Layout Knowledge Systems Lab JN 5/29/2016 Interaction Mapping Gesture to Interaction Mapping Number of Fingers: 2 – Roll Left 3 – Roll Right 4 – Zoom In 5 – Zoom Out Knowledge Systems Lab JN 5/29/2016 Gesture Recognition Demo Knowledge Systems Lab JN 5/29/2016 Exploration Solution • Doctors and radiologists: – Need to intuitively interact with the system to maximize task performance – Need to perform this interaction while being unencumbered • Solution – Can use intuitive gesturing to interact with complex, 3D data – Can interact by simply moving their hand in front of a camera, requiring no physical device manipulation Knowledge Systems Lab JN 5/29/2016 Outline • Introduction • Techniques: Segmentation, Magnification, Exploration • Solutions: – Interactive Learning – Contextual Zooming – Gesture Recognition • Conclusions and Future Work Knowledge Systems Lab JN 5/29/2016 Interactive Learning • Users can teach the computer to recognize abnormal brain tissue • They can refine the learning systems results interactively • They can save/load agents for background diagnosis on a database of medical images or to allow expert analysis in the absence of a well-paid expert Knowledge Systems Lab JN 5/29/2016 Contextual Zoom • They can zoom for precisely viewing and selecting targets/non-targets while maintaining context of the entire image • The interface facilitates task performance through interactive and customizable display of segmentation results • This system can be used with any 2D images and even with 3D datasets with some minor alterations Knowledge Systems Lab JN 5/29/2016 Gesture Recognition • Can use intuitive gesturing to interact with complex, 3D data • Can interact by simply moving their hand in front of a camera, requiring no physical device manipulation • Easily replicated and distributable • Mapping gestures to interaction is an independent stage Knowledge Systems Lab JN 5/29/2016 Gesture Recognition • Dynamic Gesture Recognition • Other interface applications include: graspable interfaces, 3D avatar / MoCap, multi-object manipulation in virtual environments, and augmented reality Knowledge Systems Lab JN 5/29/2016 Platform • Med-LIFE integration effort – Gesture Recognition has already been integrated into Med-LIFE’s Exploration tab – Contextual Zoom and Interactive learning have been combined, but not yet integrated into Med-LIFE’s learning tab • Med-LIFE will function as a single application for medical image analysis Knowledge Systems Lab JN 5/29/2016