Helbling facilitates use of AI despite limited data sets

In the development of medical applications, the engineering and consulting company Helbling uses various techniques in order to be able to implement innovations in the context of Artificial Intelligence (AI). This also makes it possible to use AI in the applications despite limited data sets, Helbling explains in a specialist article. The procedure is often necessary because the procurement of large amounts of data, for example in the form of clinical studies, is too expensive and too time-consuming, especially for small or medium-sized companies. For all techniques, specialist knowledge - including medical knowledge and an understanding of the physical processes and user interactions - is the key to successful application, the authors of the article write.

In order to be able to train an AI for medical products under these requirements, Helbling experts have attempted to reduce the model complexity and therefore also the required volume of data. “The number of parameters can be reduced by limiting the scope of the model or using extracted features instead of raw data”, the specialist article states. Using another method known as data augmentation, the experts are able to artificially enlarge the data sets. According to the experts, this involves adding slightly modified copies of existing data or generating new synthetic data. As an example, they cite a project to analyze skin cancer images. The amount of data was increased by varying images, such as by adding image noise, reducing the image contrast, changing the ambient light level or homogeneity of illumination.

The technique of transfer learning refers to the process of applying knowledge from a solved problem to a new, related problem. In this process, a subset of already trained layers of an established neural network is used and retrained. Another form of transfer learning is the combination of a small available data set with data that has been recorded by another device.

According to the authors of the article, limited data sets can also mean that unbalanced data sets exist. In such cases, it is possible to use a technique known as synthetic minority oversampling to generate new data points from an underrepresented class within a dataset.