In numerous recent studies, deep neuronal networks (DNN) have been leveraged to examine
the usefulness of artificial intelligence (AI)-based DNN for diagnostic purposes. In
essence, they have successfully proved to recapitulate state-of-the-art diagnoses
currently performed by humans.
Specifically, the use of artificial intelligence for pattern recognition showed that DNN
could categorize complex and composite data points, chiefly images, with high fidelity to
a specific pathogenic condition or disease. The majority of these studies are primarily
based on extensive training sample collections that were categorized a priori.
Subsequently, this "training" provided the necessary input to classify newly delivered
specimens into the correct subgroups, frequently even outperforming independent human
investigators. So far, these studies have thus provided the rationale for the use of DNN
in real-world diagnostics. However, the prerequisite for using DNN in a real-world
setting, where specimen sampling and analysis would need to outperform human diagnosis
prospectively, would be a blinded and prospective trial. Currently, there is a lack of
prospective data, therefore still challenging the notion that DNN can outperform
state-of-the-art human-based diagnostic algorithms. Here we want to investigate the
validity and usefulness of AI-based diagnostic capabilities prospectively in a real-world
setting.
Hematologic diagnostics heavily rely on multiple methodically distinct approaches, of
which phenotyping aberrant blood or bone marrow cells from affected patients represents a
cornerstone for all subsequent methods, such as chromosomal or molecular genetic
analyses. At the MLL, five different diagnostic pillars are required to provide
diagnostic evidence for a specific malignant blood disorder faithfully: cytomorphology
and immunophenotyping first, guiding more specific methods such as cytogenetics, FISH,
and a diversity of molecular genetic assays.
+++ Objectives +++
Phenotyping of blood cells is primarily based on two distinct challenges; (1) the
morphological appearance and abundance of specific cell types and (2) the presence of
particular lineage markers detected by flow cytometry. These two methods are critical for
each subsequent decision-making process and, thus ultimately, the final diagnosis.
Simultaneously, these two methods are ideally suited for automated analysis by DNN due to
their inherent image-based nature. This has been recently illustrated by a publication by
Marr and colleagues (Matek et al., 2019; https://doi.org/10.1038/s42256-019-0101-9)
In BELUGA, we want to investigate whether the automated analysis of blood (from
peripheral blood and bone marrow aspirates) smears and flow-cytometry-based analyses can
provide a benefit for diagnostic quality and, ultimately, patient care. Moreover, BELUGA
will provide evidence for the cooperative nature of image-based diagnostic tools for
other pillars of hematologic diagnostic decision making such as genetic and molecular
genetic characterization.
BELUGA, therefore, consists of three parts (A-C) (See Figure in the attached File). In A,
we want to train a DNN with an unprecedented collection of blood smears and
flow-cytometry-based data points collected during the course of 15 years. These samples
consist of all hematological malignancies currently identified and recognized by the
current WHO classification for hematologic malignancies. Due to the varying incidences of
these entities, the total number of training items varies from 1,000 to 20,000 for 15
years. However, we deem this discrepancy a benefit to this trial's overall aims, because
this diverse spectrum will inform us on the number of training items needed for
outperforming the state-of-the-art diagnostics in cytomorphology or flow cytometry.
In part B, we will compare the overall performance of our trained DNN prospectively to
new yet undiagnosed samples arriving at our laboratory (see the main section for
details). The superiority of DNN based categorization will be challenged based on the
pre-defined outcome parameters accuracy with respect to state-of-the-art diagnostics,
mismatch-rate, and time needed to provide a diagnostic probability.
Lastly, in C, we will investigate the effects on faster and more accurate diagnostic
power by leveraging our trained DNN to aid downstream diagnostic methodologies such as
chromosomal analysis or panel sequencing of patient samples.
