Groundbreaking results are achieved by neural networks in medical image processing. This success is reflected by a high and presumably increasing number of Deep Learning related submissions to the MICCAI conference. Despite the enormous progress in this field crucial challenges remain such as understanding the learned features, which often reside in high-dimensional space, as well as the tailor-made design and improvement of neural networks. In this tutorial, we will show how visualization can help in tackling these challenges. In traditional machine learning, features are designed instead of being learned, which is often a tedious process requiring skilled experts. We will also show how the feature design can be supported by blending visualization and data mining techniques. Causal networks supporting prognostic reasoning and discovering functional interactions are frequently being learned from complex biomedical and epidemiological data. We will show how visualization can assist an exploration and verification of the learned networks in order to, e.g., remove spurious dependencies.
Data mining systematically applies statistical and mathematical models to complex data in order to reveal patterns and trends and eventually, to infer knowledge from data. Many biomedical problems are being addressed using data mining techniques. Important challenges thereby are providing guidance to data mining laymen, e.g., physicians, in adjusting the parameters of a data mining algorithm and in interpreting its results, e.g., clusters in highdimensional space or data sub-spaces. Again, visualization has been demonstrated to be beneficial in tackling theses challenges. It assists in understanding the parameter space and cluster structure. In the tutorial, we address the blending of visualization with data mining and machine learning from a research and an applicationoriented perspective. The latter focuses on cardiac surgery planning, understand gene-structure behavior in neurosciences, tumor tissue characterization, risk factor identification in epidemiology, and clinical decision support.
The tutorial attendees will learn about:
- the difference between feature design and feature learning,
- visualization and visual analytics techniques assisting in
- (a) understanding, designing, and improving neural networks,
- (b) exploring and verifying causal networks,
- (c) understanding the parameter space of data ming methods and interpreting their results,
- typical tasks in biomedical machine learning and data mining by several application examples,
- existing frameworks and their integration in expert workflows,
- and open problems and current research trends.