View Zoom Recording
8:30-9:00 Rene Vidal (Johns Hopkins University)
Title: Machine Learning in Hematology: Reinventing the Blood Test
9:00-9:30 Marleen de Bruijne (Erasmus MC)
Title: Domain adaptation and Learning from Weak labels
9:30-10:00 Bennet Landman (Vanderbilt University)
Title: Deep Learning with Imaging Context
10:30 – 11:00 Christian Igel (University of Copenhagen)
Title: 1, 2, 3: U-Nets for Medical Data Segmentation
U-Nets are fully convolutional neural networks known for their excellent performance in image segmentation. This talk discusses U-Nets for semantic segmentation of 1-, 2- and 3-dimensional medical data, with a special emphasis on data augmentation and generalisation across clinical protocols, cohorts, and tasks. A general, lightweight system for efficient and accurate 3D image segmentation is discussed. It requires no task-specific information, almost no human interaction and is based on a fixed neural network topology and a fixed hyperparameter set, eliminating the need for model selection and its inherent tendency to cause overfitting. The talk also presents U-Time, a feed-forward network for physiological time series segmentation developed for the analysis of sleep data. U-Time maps sequential inputs of arbitrary length to sequences of class labels on a freely chosen temporal scale. The presented systems are available in open source and do not require deep learning expertise to use.
11:00 – 11:15 Junlin Yang, Nicha Dvornek, Julius Chapiro, MingDe Lin, James Duncan (Yale University)
Title: Cross-Modality Segmentation: Domain Adaptation via Disentangled Representations
Abstract: Neural network models trained on some labeled data from a certain distribution generally perform poorly on data from different distributions due to distribution mismatch, i.e. domain shifts. Unsupervised domain adaptation addresses this problem by alleviating the domain shift between the labeled source data and the unlabeled target data. We achieve cross-modality segmentation, i.e. domain adaptation between CT and MRI, via disentangled representations.
11:15 – 12:00 Mads Nielsen (Univ. of Copenhagen) & Hayit Greenspan (Tel-Aviv University)
Title: Deep Image Analysis – From research to product &OPEN DISCUSSION
This workshop will focus on the latest and greatest Deep Learning methodologies for medical image analysis. We will present and discuss the methods that have shown most robustness - over multiple tasks, across multiple institutes; along with the newest models being proposed today. In addition to network development, attention needs to be given to the training and testing methods used, to reach standardized rigorous evaluation.
Tel Aviv University
University of Copenhagen
Harvard Medical School
College of Engineering, The University of Iowa
Bram van Ginneken
Radboud University Medical Center, Nijmegen, The Netherlands
 Topics will include:
- DL methods for detection, segmentation and categorization
- Designing DL models to address more advanced medical diagnosis support tasks (e.g., staging of disease, prediction of treatment response)
- Methods for robust evaluation of DL solutions – towards high confidence solutions
- From algorithms to products: what does it take to make sure patients benefit?
- Deep learning app platforms, starting your own company, how to deal with regulatory issues