ZEISS Innovation Hub @ KIT off to a successful start

• New thesis project in machine learning
• Collaboration at many levels: Computer Vision Hackathon run with academic partner the Computer Vision for Human-Computer Interaction Lab
  

For his PhD thesis, Simon Reiß is focusing on training data and machine learning methods. Large amounts of image files are needed to perform industrial inspections and make medical diagnoses. The quality of the photographs and the subsequent image analysis are important parts of the process. A manual image assessment would not be possible due to the sheer amount of time needed. Instead, over the last few years machine learning processes have been established for automatic image analysis and processing, particularly with regard to segmentation, detection and classification.

To enable automatic image analysis, artificial intelligence processes must first be used to teach the computer where and what it should look for on the images. For example, a computer can be trained to look for pathologies and anomalies on medical images.

Current methods often have to focus on many accurately annotated image samples in order to teach the algorithm how to perform complex tasks. To learn how to automatically segment brain tumors on MRI scans, a machine learning algorithm needs to view a large number of brain MRI scans on which experts have marked the tumorous tissue. These are known as annotated training data. Since data annotations are complicated and time-consuming, this kind of training data are usually only available to a limited extent. Particularly when the algorithm is to split images into meaningful segments, marked images are needed to train the algorithms one pixel at a time.

The aim of his thesis project, titled "Image segmentation in small datasets with few annotations," is to develop a machine learning method whereby the algorithm only requires little training data or annotations to deliver a high-quality result.

We can already assess images of the eye's retina to diagnose diseases early on. To do this, experts must rely on OCT scans. These are produced using the latest image capturing procedures and are a laborious way to perform searches on the images to interpret any anomalous structures. An auxiliary system pre-processes these scans for the experts and segments any signs of disease. This way, the experts can immediately focus on any diseased tissue. Thanks to machine learning, image processing tools are already capable of doing this.

These are two areas Simon is focusing on in his thesis project. How does a system like this work with a small dataset? And: How can machine learning methods learn from less precise annotations, thus demanding much less from the experts who compile them?

Simon's PhD thesis is a joint project run by the Computer Vision for Human-Computer Interaction Lab (CV:HCI) at KIT and ZEISS. Simon Reiß benefits from his proximity to the ZEISS Innovation Hub @ KIT and, as a PhD student on the Machine Learning team at ZEISS Corporate Research, is involved with the ZEISS sites in Jena, Oberkochen and Munich. This allows him to both use and advance the machine learning and image processing methods for the ZEISS Group.

"The international ideas sharing, the wonderful tendency of industry and research institutes to share their findings and insights, and the rapid changes are what fascinate me about the Computer Vision Community", says Reiß.

The collaboration with the CV:HCI Lab at KIT goes well beyond the initial joint PhD thesis. The ZEISS Innovation Hub Computer Vision Hackathon will run from 25 to 27 September. The Laboratory acts as an academic partner. At the hackathon, ambitious and creative individuals come face to face with real-life challenges that data experts and image processing specialists at ZEISS and the KIT research group have to deal with. These range from virtually trying on glasses to helping a visually impaired person navigate a crowded area.