Past decade has seen a quantum shift in how computers perform pattern recognition tasks. A new paradigm, popularly described as machine learning, has been invented to “teach” computers how to solve problems such as classification, segmentation, and pattern recognition. The old paradigm for solving these problems in classical computer vision consisted of explicitly programming into the software, task specific features. For example, in order teach a program how to recognize human faces in an image, one would explicitly code into the software features that define a human face.
By contrast, the new paradigm of machine learning programs a computer by providing it with example data. The computer adapts its behavior by learning from these examples, much like a child learns from its surroundings. Machine learning concerns the recognition of complex patterns in observed data in order to automatically make decisions about patterns hidden in that data.
We are surrounded by examples of this new paradigm. The success of face or other object recognition programs, ability to use images instead of keywords in a search query, and the fidelity with which modern translation programs are able to transform prose from one language to another, are just a few examples of machine learning algorithms. A confluence of three trends —namely, massive computation power, easy availability of “big” data (e.g., through social media), and advent of a new computational architecture called deep neural networks— has made such rapid and widespread adoption of machine learning feasible.
Just as this paradigm has fundamentally altered the discourse on the social media and the internet, machine learning will fundamental alter nearly all aspects of medical practice of future. Medical systems, e.g., CT and MRI scanners, ECG machines, EEG and other physiologic monitors, produce huge amounts of data that often contain abundant information. Only a fraction of this information is important for the diagnosis. The new paradigm of machine learning raises several deep and incisive questions.
Can machine learning automate and help to extract relevant information from vast amounts of medical data?
- What are the various paradigms for organizing and extracting relevant information?
- Is it possible to teach a computer to make automated decisions after a session of example-driven learning?
- More fundamentally, can we make a medical diagnostician obsolete, relegating the diagnostic exercise to a computer?
These are the central questions that we will be addressing in this course.
This course will help a student acquire knowledge, skills and insight in machine learning in the domain of medical imaging and sensor data. The course will begin with an introduction to the classical techniques in classification, supervised and unsupervised learning, and regression. This will be followed by the newly developed field of machine learning. While the concepts are general, the focus will be on medical applications of these concepts.
Examples of polyp, ischemic stroke, retrained surgical hardware (sponge) after surgery, and subarachnoid hemorrhage. Can a computer learn to automatically detect them after looking at thousands of such examples?
The course addresses the machine learning and deep learning paradigm for classification and pattern recognition in vast amounts of data. Examples: How can we program a system in such a way that anomalous frames in an endoscopic video stream are detected automatically? How can we delineate an anatomical structure? In the machine learning paradigm, the solution is found by providing the system with a large number of examples from which, in a training stage, the system tries to generalize. Once trained, the system is then ready to be used on any individual image in the operational stage.
This course is an introduction to classical machine learning and the new paradigms such as convolutional neural nets for deep learning. The topics that will be covered include:
- Architectures for classification and regression
- Overfitting, underfitting, and the generalization gap
- Strategies for learning such as convolutional neural networks
- Hardware and software platforms for machine learning
- The medical context and significance of machine learning
Each day of this 5-day course is organized as follows:
- Introduction to the theoretical topics of the day
- Clinical context and clinical relevance of the topic covered
- Matlab hands-on exercise illustrating the theoretical concepts with clinical examples
- Re-cap, closing remarks, and closure of the day
- Assessment/Examination for the day
The students will be expected to program extensively by themselves. This will enable them to quickly apply their newfound knowledge and learn the material from a practical stand point.
Day 1: Introduction to Machine Learning paradigms and their significance to medicine.
Two basic classification algorithms will be outlined. Hands-on experience will be offered in which these classifiers are trained with data from a relevant clinical problem.
Day 2: Learning strategies for more advanced classification algorithms.
More advanced classifiers will be introduced, such as the Support Vector Machine (SVM), and the feedforward neural network with backpropagation. The pitfalls which might occur during the training of these classifiers will be elucidated with a clinical example. In this exercise, also strategies for countermeasures will be demonstrated.
Day 3: Feature generation and reduction.
Medical images are 3D data volumes that carry rich information about anatomical and functional structures and possible diseases. However, the 3D image data are voluminous, low-level representations of information. Feature generation, such as in radiomics, reduces the data size by grasping the local statistics of the image in so-called features. Techniques for a further, application dependent reduction of number of features is the topic of day 3.
Day 4 and 5: Deep Learning in Medicine
Deep learning networks, and specifically the convolutional neural network, will be introduced. Their clinical significance, based on an historical overview, will be presented, together with the implementation aspects. Participants will be offered hands-on training with a simple convolutional neural network.
This course is intended for PhD students and young scientists with a background in Biomedical Technology, Technical Medicine, and related fields. It is primarily intended for students who have a need for more insight in machine learning with application to medical image data.
The course assumes a background in programming. In addition to the medical relevance of the topics being covered, the exposition to the theoretical concepts will be made more concrete by a set of well selected programming exercises.
Familiarity and experience with Matlab and with vector-matrix calculus is a prerequisite for this course.
The course is organized by:
- Dr Ferdinand (Ferdi) van der Heijden, PhD – University of Twente
- Prof Rajiv (Raj) Gupta, MD, PhD – Massachusetts General Hospital and Harvard Med School
- Dr Beril Sirmaçek, PhD – University of Twente
- Can (John) O. Tan, PhD – Spaulding Rehab Hospital, MGH and Harvard University
- Floortje Jolink – Masters Student, Medical Imaging and Interventions, Technical Medicine, University of Twente
- Rick Bergmans – Masters Student, Medical Imaging and Interventions, Technical Medicine, University of Twente
under the auspices of the Technical Medicine department of the University of Twente in collaboration with the Medical Imaging group of the MIRA Institute for Biomedical Technology and Technical Medicine.
- The course is part of the regular UT curriculum and welcomes attendance from external participants
- The course will run from 28 May – 1 June 2018
For UT Master students
- Attending the course is free of charge for UT master students
- The course will represent 1.5 EC
- This ML course is part of the optional course “3D Computer Vision for Medical Applications” (5 EC). If you enrol yourself for that course, then you are automatically also enrolled for this ML course
- Enroll yourself in OSIRIS for either the 5 EC (201500553) or the 1.5 EC (201500583) course