August 22, 2023 | Over the next three years, researchers from across five universities in Germany will be using “swarm learning” on retrospective tissue samples to gain insights on the tumor biology of hereditary colorectal cancer (CRC). The decentralized artificial intelligence (AI) approach combines edge computing and blockchain technology, enabling scientists to “bring the algorithm to the data” and in sufficient quantity to bolster the model’s performance, according to Sebastian Försch, Ph.D., assistant physician and research assistant at the Institute of Pathology at the University Medical Center in Mainz, Germany.
The project, which launches this year, is known as the DECADE study—short for decentralized artificial intelligence for diagnosis, prognostication, and response prediction in colorectal cancer—and is funded by the nonprofit German Cancer Aid. Collaboration partners come from university hospitals in Bonn, Dresden, Düsseldorf, Heidelberg, and Mainz.
Most of the partners are already part of a loose network of research groups coupling gastrointestinal oncology with computational biomedicine, Försch says. The concept was the brainchild of internist and scientist Jakob N. Kather, M.D., professor of clinical artificial intelligence at the Else Kröner Fresenius Centre for Digital Health at the TU Dresden and the University Hospital Dresden, and leading experts in related disciplines (e.g., hereditary CRC and computational pathology) were easily recruited.
When it comes to developing AI models for healthcare applications, the key difficulties are coming up with enough data to train the model and the fact that sensitive medical information can’t be shared, says Försch. This is enormously challenging in Germany, where the rules pertaining to the security and protection of personal data are particularly strict.
The workaround is to instead share the parameters of the trained model, he adds, which is what swarm learning is all about. “The blockchain technology helps to make sure that only authorized members of the swarm have access and that everything is documented and secured. You can even encrypt the parameters, to add an extra level of security.”
The AI models to be used here are primarily algorithms from the field of computer vision, optimized for the analysis of histopathological data, Försch continues. One of the earliest published studies on the technique appeared in Nature in 2021 (DOI: 10.1038/s41586-021-03583-3), and Kather and his team were the first to apply the concept to histopathology the following year, as reported in Nature Medicine (DOI: 10.1038/s41591-022-01768-5).
Improving the diagnosing and subtyping of CRC and predicting disease progression are the goals of the DECADE study, whose underlying structure is much like that of other retrospective biomarker trials. One major difference is that investigators aren’t looking at a particular molecule or pathway; rather, “swarm learning/AI from histopathology is the biomarker,” Försch says.
The DECADE study also has several other unique aspects, including the fact that it is focused not only on sporadic CRC but also the hereditary types, given the particularly strong medical need to improve disease detection and treatment in this subgroup of patients, says Försch. Additionally, the research team is trying to mirror the entire course of the disease with subprojects variably looking at early detection and characterization of colorectal precursor lesions, primary tumors, and distant metastasis when the tumor has already spread.
The plan is to gather retrospective tissue material as well as validation samples from newly diagnosed patients, for a total enrollment of hundreds if not thousands of patients, he reports.
This is the first CRC trial to use swarm learning, says Försch, and could serve as a template for any AI system in the healthcare sector. Several other collaborative projects are underway to investigate use of the approach for other applications, among them the so-called SWAG project that combines swarm learning and generative AI in kidney cancer and is funded by the Federal Ministry of Education and Research in Germany.
“AI applications have the potential to transform healthcare and modern medicine and thereby improve diagnostics and treatment options for patients with cancer and other diseases substantially,” he notes. “However, it is really important not to fall victim to some type of AI-hype, but to thoroughly investigate the advantages and disadvantages of these types of applications for patients, doctors, and other healthcare providers,” he adds, in pointing to the advantages of the translational, interdisciplinary approach planned in the DECADE trial.
Bowel cancer is one of the most common and deadly types of cancer in Germany. About 58,000 people per year are diagnosed with the disease, which is easily curable but only if detected early, Försch says.