Download as PDF, PPTX
Uploaded byCodeOps Technologies LLP
Biomedical Signal and Image Analytics using MATLAB
The document discusses the use of MATLAB for biomedical signal and image analytics, emphasizing its accessibility for engineers and scientists. It showcases various applications, including algorithm development for motion analysis, respiratory monitoring, and heart sound classification, illustrating reduced development times and improved algorithm performance. The document also highlights MATLAB's capability in deep learning and data analytics workflows, enabling users to efficiently create and deploy sophisticated algorithms.
More Related Content
Similar to Biomedical Signal and Image Analytics using MATLAB
More from CodeOps Technologies LLP
Biomedical Signal and Image Analytics using MATLAB
- 1. 1© 2018 TheMathWorks, Inc. Biomedical Signal and Image Analytics Using MATLAB Dr. Amod Anandkumar Senior Team Lead – Signal Processing & Communications Application Engineering Group @_Dr_Amod
- 2. 2 What they say •Expand university programs • Train existing analysts
- 3. 3 The need fordata scientists Domain expertise Coding and integration skills Statistical and mathematical knowledge
- 4. 4 MATLAB lets yoube your own data scientist MATLAB is Designed and Documented to be Easy for Engineers and Scientists to Use “As a process engineer I had no experience with neural networks or machine learning. I worked through the MATLAB examples to find the best machine learning functions for generating virtual metrology. I couldn’t have done this in C or Python—it would’ve taken too long to find, validate, and integrate the right packages.” Emil Schmitt-Weaver, ASML “MATLAB enabled us, as geologists, to use our expertise in predictive frameworks, analytics, and analog matching to implement algorithms that are unique in our industry. With the help of MathWorks consultants, we then deployed those algorithms as an easy-to-use application to our colleagues worldwide.” Nick Howes, Shell
- 5. 5 Developing Motion AnalysisAlgorithms for Medical, Sports, and Occupational Safety Applications Challenge Accurately measure and track motion for medical, sports, and occupational safety applications Solution Develop motion-tracking algorithms in MATLAB and deploy them to a wearable, wireless motion analysis device with MATLAB Coder Results ▪ Time required to translate algorithms to production C code reduced from a month to a few days ▪ Development time reduced by 50% ▪ Defect-free code generated “With MATLAB and MATLAB Coder, we can respond quickly to customer requests. In a recent release, we updated two algorithm modules, created two new ones, and delivered all four with a confidence that would have been impossible using our old workflow.” Edgar Charry dorsaVi Link to article dorsaVi sensors in use in a running test.
- 6. 6 Challenge Develop and implementan acoustic respiratory monitoring system for wheeze detection and asthma management Solution Develop algorithms for detecting wheeze and ambient noise in MATLAB, and use MATLAB Coder to generate code from the algorithms for mobile devices and a web server Results ▪ Manual coding effort reduced ▪ Algorithm development iterations accelerated ▪ Code maintenance overhead reduced Respiri Develops Mobile App and Server Software for Wheeze Detection and Asthma Management “MATLAB enables us to rapidly develop, debug, and test sound- processing algorithms, and MATLAB Coder simplifies the process of implementing those algorithms in C. There’s no other environment or programming language that we could use to produce similar results in the same amount of time.” Mark Mulvey iSonea The AirSonea device, which connects to an asthma patient’s smartphone and communicates with wheeze analysis algorithms on iSonea’s server. Link to user story
- 7. 7 Data Analytics Workflow IntegratedAnalytic Systems Desktop Apps Enterprise Scale Systems Embedded Devices and Hardware Files Databases Sensors Raw Data Models and Algorithms Model Creation e.g. Machine Learning Model Validation Parameter Optimization Preprocessed Data Working with Messy Data Data Reduction/ Transformation Feature Extraction
- 8. 10 Today’s Examples ▪ Heartsound classification using machine learning ▪ MR image segmentation using deep learning ▪ Using visualizations for understanding CNN behaviour
- 9. 11 Heart Sound Classification– An Example ▪ Motivation – Heart sounds are complex signals – Require trained clinicians for diagnosis ▪ Objective – Develop algorithm for automatic classification of heart sounds ▪ Dataset – Training: 3240 recordings – Validation: 301 recordings – Recording Duration: 5 to 120 seconds – Labels: Normal or Abnormal Heart Sound Recording Classification Algorithm Normal Abnormal PhysioNet/CiC Challenge 2016 View Webinar
- 10. 12 Hyperparameter Tuning andFeature Selection Bayesian Optimization ▪ Tune hyperparameters of machine learning algorithms automatically ▪ Use bayesopt function for complete control. – Define objective function to be optimized. – Optimize under constraints you define. Neighbourhood Component Feature Selection ▪ Feature selection for classification (fscnca) or regression (fsrnca) ▪ Suitable for use before any classification or regression method ▪ Produces weights that go to 0 for irrelevant features ▪ Regularization parameter λ can be selected by cross-validation
- 11. 13 Wavelets for SignalProcessing and Feature Extraction ▪ Time-Scale Analysis is more robust and powerful ▪ Signal Morphology & Features can be extracted ▪ Excellent for pre- processing and feature engineering on signals Detrending ECG Signal Continuous Wavelet Transform of an ECG Signal
- 12. 14 Statistics & MachineLearning ▪ Use apps to interactively explore data, choose algorithms to train and validate models, and compare results – Classification Learner app – Regression Learner app ▪ Apply algorithms to out-of-memory data using tall arrays ▪ Generate C code for predictive models (requires MATLAB Coder) MATLAB makes machine learning easy and accessible for everyone, even if you’re not an expert “I would have never attempted machine learning if this app was not available.”
- 13. 15 Today’s Examples ▪ Heartsound classification using machine learning ▪ MR image segmentation using deep learning ▪ Using visualizations for understanding network behaviour
- 14. 17 Perform Object Detectionand Recognition using Deep Learning Deep Learning ▪ Convolutional Neural Networks ▪ Directed Acyclic Graph Networks ▪ Autoencoders ▪ Long Short-Term Memory Networks Computer Vision System Toolbox Neural Network Toolbox Parallel Computing Toolbox Object Detection ▪ Regions with CNN (R-CNN) ▪ Fast R-CNN ▪ Faster R-CNN Deep Learning … 𝟗𝟓% 𝟑% 𝟐% Truck Car ✓ Bicycle Learned features End-to-end learning Feature learning + Classification
- 15. 18 Deep Learning Challenges ▪Accessing reference models from research ▪ Difficult to understand network behavior ▪ Tuning hyperparameters and refining architectures ▪ Training takes hours-days ▪ Labeling large sets of images ▪ Real world systems use more than deep learning – Deep learning frameworks do not include “classical” computer vision ▪ Deployment on embedded hardware ▪ Not a deep learning expert
- 16. 19 Convolutional Neural Network Convolution+ ReLU PoolingInput Convolution + ReLU Pooling … … Flatten Fully Connected Softmax dog … cat … … Feature Learning Classification goldfish✓
- 17. 20 Classification vs/ Regressionv/s Detection v/s Semantic Segmentation HIGHWAY_SCENE Classification Car Car CarDetection Lane Lane Regression Semantic Segmentation
- 18. 21 Cardiac MR LeftVentricle Segmentation – An Example ▪ Motivation – Cardiac MR images are very complex – Require trained clinicians for manual segmentation ▪ Objective – Develop algorithm for automatic segmentation of left ventricle ▪ Dataset – Training Data: 260 images – Pixel Labels: LV or background LV Segmentation Challenge
- 19. 22 Access Large Setsof Images Handle Large Sets of Images Easily manage large sets of images - Single line of code to access images - Operates on disk, database, big-data file system imageData = imageDataStore(‘vehicles’) Easily manage large sets of images - Single line of code to access images - Operates on disk, database, big-data file system Organize Images in Folders (~ 10,000 images , 5 folders)
- 20. 23 Augment Training Images Rotation Reflection Scaling Shearing Translation Colourpre-processing Resize / Random crop / Centre crop
- 21. 24 Training Deep NeuralNetworks trainingOptions ▪ Plot training metrics – Training accuracy, smoothed training accuracy, validation accuracy – Training loss, smoothed training loss, and validation loss ▪ Debug training – Stop and check current state – Save / load checkpoint networks – Custom output function (stopping condition, visualization, etc.) ▪ Bayesian optimization for hyperparameter tuning Learn More
- 22. 25 Training in MATLABis fast MATLAB is more than 4x faster than TensorFlow AlexNet CNN architecture trained on the ImageNet dataset, using batch size of 32, on a Windows 10 desktop with single NVIDIA GPU (Titan Xp). TensorFlow version 1.2.0.
- 23. 26 Alexnet Inference onIntel CPUs MATLAB (R2017b Release 2) TensorFlow MXNet Caffe2
- 24. 27 Scaling Training onCPU, GPU, Multi-GPU and Clusters More GPUs MoreCPUs
- 25. 28 Semantic Segmentation Support for ▪Fully convolutional networks ▪ SegNets ▪ U-Nets and other directed acyclic graph (DAG) networks Learn More Manage connections ▪ addLayers, removeLayers ▪ connectLayers, disconnectLayers Manage label data and evaluate performance ▪ pixelLabelDatastore ▪ evaluateSemanticSegmentation ▪ jaccard,dice,bfscore
- 26. 29 Perform Transfer Learningwith 10 lines of MATLAB Code Load Pre-Trained Network Replace Final Layers Train Network on New Data Load New Training Data Example Available Here
- 27. 30 Import Pre-Trained Modelsand Network Architectures Pretrained Models ▪ AlexNet ▪ VGG-16 ▪ VGG-19 ▪ GoogLeNet ▪ Resnet50 ▪ InceptionV3 ▪ Resnet101 Import Models from Frameworks ▪ Caffe Model Importer (including Caffe Model Zoo) – importCaffeLayers – importCaffeNetwork ▪ TensorFlow-Keras Model Importer – importKerasLayers – importKerasNetwork Download from within MATLAB net = alexnet; net = vgg16; net = vgg19; net = googlenet; net = resnet50; net = inceptionv3; net = resnet101;
- 28. 31 Visualizations for UnderstandingNetwork Behavior ▪ Custom visualizations – Example: Class Activation Maps Filters … Activations Deep Dream http://cnnlocalization.csail.mit.edu/Zhou_Learning_Deep_Features_CVPR_2016_paper.pdf
- 29. 32 How can Icreate ground truth? Specify ground truth to train detector Object Detector Semantic Segmentation Network TrainingGround Truth Images
- 30. 33 Accelerating Ground TruthLabeling Image Labeler App – Sequence of images - Bounding boxes, image pixels Use automation algorithms to accelerate labeling
- 31. 34 Generate CUDA codeand deploy CNNs with GPU Coder Alexnet Vehicle Detection People detection Lane detection ~30 Fps (Tegra X1) ~66 Fps (Tegra X1) ~20 Fps (K40c) ~130 Fps (K40c) New Product: GPU Coder ▪ Portable and readable CUDA code generation ▪ Eliminate manual coding errors ▪ Maintain Single “Truth” ▪ Stay within MATLAB at a higher level
- 32. 35 Alexnet Inference onNVIDIA Titan Xp MATLAB GPU Coder (R2017b) TensorFlow (1.2.0) Caffe2 (0.8.1) Framespersecond Batch Size CPU Intel(R) Xeon(R) CPU E5-1650 v3 @ 3.50GHz GPU Pascal Titan Xp cuDNN v5 Testing platform MXNet (0.10) MATLAB (R2017b) 2x 7x5x
- 33. 37 MATLAB Beyond theDesktop Share Applications Integrate with Web & Enterprise Scale Computation Scale Data
- 34. 38 Implement and PrototypeAlgorithms in Hardware VHDL, Verilog ASIC FPGA HDL Coder Programmable SoC DSP / MCU Embedded Coder C, C++ ROS Node Robotics System Toolbox CUDA GPU Coder
- 35. 39 Resources to learnand get started mathworks.com/machine-learning mathworks.com/deep-learning mathworks.com/solutions/medical-devices.html
- 36. 40 Deep Learning Onramp FreeIntroductory Course Available Here
- 37.