Heidelberger BV: Challenges of Machine Learning - from Annotation to Productization
The productization of machine learning methods in medical technology, such as the use of deep neural networks (DNNs) for object recognition or segmentation of medical image data, requires long-term planning and poses new challenges for many companies. Training DNNs requires a sufficiently large amount of quality-assured and already annotated data, which must be processed in a GDPR-compliant manner. Designing and training of problem-specific DNNs for large volume data from CT/DVT or MR scanners is another challenge. These quickly become very memory-hungry and place high demands on the hardware, which must be taken into account during productization.
Among other things, this also requires an adaptation of already trained DNNs to the hardware requirements of the product. For product integration, a suitable DNN framework should be selected that is scalable and easy to maintain with future product extensions. However, these in particular are very new technologies that are rapidly evolving. Medical devices in the EU are subject to the Medical Device Regulation (MDR) after market launch. This requires monitoring of product performance and any necessary improvements to the product. This places
requirements on the DNN product design in order to improve them efficiently in the future.
The presentation will give examples and an experience report about the entire process chain of a productization in medical technology. Examples of a quality-assured annotation process are shown and semi-automatic annotation tools are presented. Additionally, it gives an insight into the design and training of DNNs. Tools for the subsequent productization of already trained DNNs will be presented and selected DNN frameworks are explained with respect to product integration.