Applying machine learning for the comprehensive gathering and analysis of biosignals
Because vital biosignals are now being continually monitored on a daily basis thanks to wearables and smart integrated sensor technology, we have access to large quantities of medical data that are extremely useful in diagnosis and treatment. When applied correctly, deep leaning techniques help find patterns in biosignals and allow us to accurately classify them according to various clinical categories (benign/malignant etc.). The information we glean from this process regarding the way in which biosignals interact with one another forms the basis of important medical decisions and preventative treatment.
Applying machine learning and deep learning for user-focused human-machine interaction
Another key topic of research is the development of methods and systems for machine learning. Our focus in on (real-time) computer vision applications (including facial analysis) that deliver high performance and detection rates even in uncontrolled environments. Another of our goals is to make our technology available across a wide range of platforms and in the embedded sector – which calls for efficient implementation and intelligent algorithm design. Our topics of research encompass state-of-the art techniques, such as those from deep learning, as well as approaches based on classic machine learning. Our technology is implemented in a range of applications, including market research, robotics, stress detection and the automotive sector.
Endoscopes have been used by physicians for over a century to diagnose and treat diseases and injuries affecting the body’s internal tissues and organs. But endoscopic examinations are still a purely visual technique and must often be complemented by biopsies to provide a full diagnosis.
The interpretation of endoscopic images depends to a large extent on the physician’s prior experience and is thus quite subjective. Fraunhofer IIS is investigating and developing methods for characterizing and classifying tissue damage and detecting suspect lesions, such as intestinal polyps, based on machine learning and using, for instance, specific color and texture features. The methods are intended for applications in diagnostic endoscopy.
Digital pathology and computer-assisted microscopy
The trend toward precision medicine calls for ever more nuanced diagnoses that take a multitude of histologically defined biomarkers and a growing number of prognostic factors into account. The application of machine learning methods in digital pathology holds much promise as an efficient means of mastering these challenges. The pathologist is relieved of routine diagnostic tasks that can now be automated, such as detecting and analyzing relevant cells and areas of tissue, and the results can be visualized in context. In research, deep learning methods assist in identifying and validating new prognostic markers and therapeutic targets.
Our researchers are well versed not only in conventional machine learning and image analysis methods, but also in the new, often superior, deep learning methods, and possess many years of experience in the development of applications to support computer-assisted diagnosis. These applications range from microscope and scanner control functions and automated analysis routines to the presentation of reproducible results in an easily comprehensible visual format.