Jump to contentJump to search

Can pigeons classify pathology images?

Yes, they apparently can, according to this paper published in PLOS ONE in 2015. But as Jun.-Prof. Dr Jakob Nikolas Kather pointed out in his recent HeiCADLecture on ‘Predicting genetic alterations in solid tumors from histological images with deep learning’, the image classification itself might be easy due to the excellent performance of deep learning tools while useful medical application of it is actually hard. Some examples of what the classification of H&E-stained histopathology images can be used for are differentiation of whether a tumour is present or not, tumour genotype, treatment response and overall survival of a patient. Jun.-Prof. Kather presented work conducted in his group ranging from applications in gastroenterology to urology (those works and publicly available datasets can be found on the group’s webpage kather.ai).

In the second part of his talk, he gave an outlook of what the next steps involve: new approaches such as multiple instance learning need to be further established, application can be extended to other diseases and imaging modalities, and generative methods might be explored to obtain synthetic data. Moreover, privacy and data security continue to play an important role. Finally, the human factor is crucial. Deciding what hardware to use, implementing algorithms correctly and efficiently as well as acquiring, processing and analysing the data - all of this requires an interdisciplinary team of researchers.

Jun.-Prof. Kather believes that every medical practitioner is going to be confronted with AI technology in the future, so it is important to know about the basics, know about potential biases and be able to make judgements.

Autor/in: Dr. Joana Grah
Responsible for the content: