Discussion
Our study showed that using a deep learning approach to analyze MIR
imaging data from unstained histological slides can distinguish between
benign and malignant lymphoid tissue and aid in classifying types of
lymphomas. MIR imaging on tissue sections results in large and complex
data sets that must be analyzed and interpreted. Focussing on possible
use in routine diagnostics, the data acquisition and the subsequent
analysis must take place within a narrow time window to offer an
advantage over standard pathological diagnostics or assist the
pathologist during the diagnostic procedure. A deep learning approach
using NNs, offers such a possibility. Our NN, once trained, analyzed a
MIR image in less than 25 ms. However, training NNs using a deep
learning approach takes considerable time and requires graphical
processing units (GPUs) and high-performance machines to process digital
images [28, 30]. The training, however has to be performed only
once, and the NN can then be used to classify IR data without further
training. This even allows for real-time spectrum analysis.
NNs have already been implemented in medicine and used in various
studies in the fields of radiology, cardiology, neurology, and pathology
[30-34]. But several caveats prevent such techniques from being
widely used. One of the most important is undoubtedly the ”black box”
character of such analyses [29, 30]. The decisions of such
algorithms are not easy for the human user to understand and interpret.
Therefore, a certain scepticism about such approaches is understandable,
especially in medicine, where comprehensible decisions with sometimes
severe consequences for patients must be made and justified [28,
29].
Another problem is the need for large datasets to train NNs for a
specific problem [28, 30]. Small datasets result in poor performance
of NNs, while on the other hand, one has to bear in mind not to
”overfit” a NN [29, 30]. In the case of pathological diagnoses,
obtaining such datasets is a significant problem, especially regarding
rare diseases. But one has also to consider the extreme variability of
histopathological patterns derived from various tissue types, building
up organs [28]. Additionally, data quality is essential, especially
when dealing with pixel-based data (images) in pathology, where
artefacts might pose serious problems [28, 35, 36]. In fact,
concerning MIR imaging, there are currently no available datasets that
could be used, and it is also questionable if there will be any in the
future. Overall, it is doubtful that any deep learning approach using
NNs, combined with MIR microscopic imaging, will replace pathologists in
the foreseeable future. It is more likely, that deep learning-driven
approaches will be used to assist a pathologist during the diagnostic
procedure [37].
Finally, besides technical questions and dataset availability, ethical
and legal questions are also associated with the use of deep learning in
decision-making processes in pathology. These fundamental questions
range from concern about data privacy to the question of responsibility
for a wrong decision based on a deep learning approach [29].
However, when there are fewer and fewer pathologists, deep learning
techniques may assist as a diagnostic tool to support the pathologist in
stratifying patients, identifying urgent cases, and thus better
directing the routine workflow.
Finally, the financial side of introducing such technologies must also
be considered. The financial pressure on pathology laboratories is
already a challenge because of the increasing digitization and
subsequent data storage [28]. The acquisition of access to the
appropriate hard- and software, such as GPU clusters, as a must to train
deep learning algorithms in practice, could fail due to a lack of
funding [28, 38].
Our results demonstrate that a diagnosis and subtyping of human lymphoma
could be streamlined based on MIR microscopy and a deep learning
approach for data analysis. This might be a complementary pathway for a
quick preliminary assessment of the type and aggressiveness of the
disease and could probably help in advance to identify urgent diagnoses
and, given the increasing shortage of pathologists, to prioritize these
patients in the routine workflow.