Program Information

Deep Learning and Applications in Medical Imaging

B Sahiner

L Hadjiiski

M Giger

R Summers

B Sahiner¹*, L Hadjiiski²*, M Giger³*, R Summers⁴*, (1) US Food and Drug Administration, Silver Spring, MD, (2) University of Michigan, Ann Arbor, MI, (3) University Chicago, Chicago, IL, (4) Warren Grant Magnuson Clinical Center, Bethesda , MD

Presentations

1:45 PM : Deep learning: Principles, achievements and future potential in medical imaging - B Sahiner, Presenting Author
2:15 PM : Can deep learning help in cancer diagnosis and treatment? - L Hadjiiski, Presenting Author
2:45 PM : Role of deep learning at various stages of quantitative image analysis for disease assessment - M Giger, Presenting Author
3:15 PM : The impact of deep learning on radiology - R Summers, Presenting Author

MO-DE-601-0 (Monday, July 31, 2017) 1:45 PM - 3:45 PM Room: 601

Deep learning techniques use massive and multi-layer networks of artificial neurons for complex tasks that require highly nonlinear behavior and large learning capacity. Similar to the networks in the visual and auditory cortex of the human brain, deep learning systems can automatically discover compact hierarchical feature representations using large datasets of labeled or unlabeled training samples. Systems using deep learning have been successfully applied to a wide range of problems in object recognition, speech recognition, and natural language processing. In image analysis, a particular type of deep learning known as deep convolutional neural network (CNN) has been shown to outperform traditional machine learning techniques across an array of tasks when the dataset size is large. Application areas in medical imaging include feature extraction, segmentation, image reconstruction, and computer-aided diagnosis.

In this session, we will first review the basic building blocks of artificial neural networks (ANNs). We will review how ANNs are trained, and how the ability of a trained network to perform a task generalizes from the training cases to independent test cases. We will compare CNNs with more traditional ANNs for imaging applications, and introduce some of the specialized techniques typically used in the design of high-performance deep CNNs. We will compare popular deep learning software frameworks and libraries, many of which can facilitate the obstacles related to architecture design, training and efficiency for newcomers to the field. We will discuss applications of deep learning to medical data sets, including feature extraction, segmentation, quantitation, and classification, with specific examples from our current research. We will cover particular pitfalls and roadblocks encountered in medical imaging such as limited dataset set size, and examine potential remedies, such as data augmentation and transfer learning. We will conclude by summarizing current challenges and future opportunities.

Learning Objectives:
1. Understand basic building blocks and training methodology for deep learning systems and convolutional neural networks.
2. Learn about publicly available software, as well as architecture, hardware and training speed issues.
3. Understand applications of deep learning in medical image feature extraction, segmentation, classification and computer-aided diagnosis with specific examples.
4. Learn about current opportunities and challenges in medical imaging.

Handouts

• 127-35326-418554-126897.pdf (B Sahiner)

• 127-35327-418554-126371.pdf (L Hadjiiski)

• 127-35328-418554-126330.pdf (M Giger)

• 127-35329-418554-127442.pdf (R Summers)

Contact Email: