BELUGA - A Prospective, Registered Study on the Use of Artificial Intelligence (AI) in Hematology

Even before the start of the COVID-19 pandemic, we saw how strongly digitalization is shaping our private and professional lives and will continue to do so in the future. This also applies to innovations for daily work in the paperless Munich Leukemia Laboratory (MLL), such as the complete digitalization of all medical findings in a constantly improving digital infrastructure. However, AI will mark a new era for medical diagnostic services as well. 

In the last few months, many scientific publications have addressed this issue, examining the use of artificial intelligence in numerous clinical applications. In these studies, Deep Neural Networks (DNNs) are often used to test the application of machine learning for diagnostic purposes. Ranging from automated fundus examinations, to the detection of tumor foci in tissue sections and the detection of COVID-19-disease from CT scans, these systems are for the most part not only equivalent to human examiners in terms of accuracy and speed, but actually increasingly becoming superior. However, most of these studies are retrospective: A large collection of pre-annotated data (or images) is used as a training cohort to have new, previously unannotated data classified by the network.

We are now starting the prospective part of the MLL-initiated BELUGA study (Better Leukemia Diagnostics Through AI; Clinicaltrials.gov, NCT04466059). We will investigate to what extent AI-based diagnostic work-up is on a par with, or superior to, the conventional gold standard.

Our AI approaches draw on a collection of currently over 600,000 digitized blood cells and over 300,000 digital immunophenotypic findings, which have undergone state-of-the-art interpretation and annotation at the MLL. The DNNs trained with the MLL data are now available as applications. As part of the BELUGA study, these DNNs will be used in parallel to the current gold standard of routine laboratory practices and workflows over a period of one year to prospectively evaluate new incoming patient samples. AI results will remain blinded. The two methods will then be compared at specified time points in relation to defined primary endpoints (sensitivity/specificity of the diagnoses, time to diagnosis). Other secondary endpoints will include step-by-step diagnostics focusing on molecular genetics and chromosome analyses.

For the first time, the potential of AI-guided diagnostic strategy is being investigated in a prospective study using the example of cytomorphology and immunophenotyping. The newly collected data will be used to continuously train the algorithm. It would be exciting if, 200 years after the birth of Rudolf Virchow, the first person to describe “white” blood, the approach of the BELUGA (Russian for "white") study described here indeed helped AI-based diagnostic strategy find its way into hematologists' daily clinical routine.