New Machine Learning in Medical Imaging Journal Club Adam Alessio, PhD, Radiology David Haynor, PhD, MD, Radiology
Machine Learning in Medical Imaging Goals: Introductions Brief Review of Terminology (review Erickson 2017 Radiographics ) Discuss Scope of Journal Club Topical areas Format of meetings Initial list of possible journal articles Potential other offerings if of interest
Introductions Name, Department, Familiarity with Machine Learning Methods: 1 None 2 Some Familiarity (aware of basic concepts, no hands-on experience) 3 Familiar (some hands-on, taken intro class, read some lit) 4 Very Familiar (published papers on these topics, active ongoing projects) 5 Expert (developed new algorithms, wrote book chapter, numerous papers)
Labeled Data: Set of examples with correct answer Erickson et al, Machine Learning for Medical Imaging. Radiographics, 2017 Machine Learning: Method to learn from a set of data, then apply methodology to make a prediction Classification: Task of assigning a class or label to data (binary or more) Training: Phase during which algorithm system has labeled data with answers Labeled Data: Set of examples with correct answer Training Set: Validation Set: Testing Set: For testing performance of a trained system Node, Layer, Weights : Components of algorithm (usually nodes and layers predefined and training process finds optimal weights)
Supervised vs Unsupervised Methods: Erickson et al, Machine Learning for Medical Imaging. Radiographics, 2017 Supervised vs Unsupervised Methods: Supervised: learning from labeled/annotated data Unsupervised: Learning from unlabeled data (no knowledge of groups present in training data) Reinforcement Learning: Start with training data with labels. Then system improves with unlabeled data. -------- Feature Computation/Extraction – extraction of info to make decisions Feature Selection – Selecting subset of features
Conventional Machine Learning: Erickson et al, Machine Learning for Medical Imaging. Radiographics, 2017 Conventional Machine Learning: Neural Nets K-Nearest Neighbors Support Vector Machines Decision Trees Naive Bayes (assumes features statistically independent) Deep Learning: multiple layers of nonlinear processing units (>20 layers) and the supervised or unsupervised learning of feature representations in each layer Examples: deep neural networks, convolutional neural networks, deep belief networks and recurrent neural networks
Other Items: Schedule?, Other Offerings?, Concerns? Machine Learning in Medical Imaging Suggested Scope For Journal Club Discussions: Start out with deep neural networks and then move on to other applications and technologies later as time/interest warrant, and our scope may include biomedical (non-radiology) imaging applications as well. Other Items: Schedule?, Other Offerings?, Concerns?
Figueiredo. Nature biomed eng. 2017
Esteva Nature. 2017
Esteva Nature. 2017
https://www.kaggle.com/c/noaa-fisheries-steller-sea-lion-population-count
https://www.kaggle.com/
Training Set Has 3500-5000 images
Gulshan JAMA. 2016
Gulshan JAMA. 2016