Page 63 - AIH-1-2

Artificial Intelligence in Health AI in AD – Diagnosis and monitoring

is neurological imaging. Positron emission tomography including cardiology, neurology, and gastrointestinal
(PET) and functional magnetic resonance imaging research, in addition to dermatology. 34
(fMRI) have been used to measure brain activity during
experimentally induced itching episodes, offering insights 10. Limitations of AI in dermatology and
into treatments for acute itching. In a previous study, possible solutions
73
eight patients with AD and six healthy controls underwent At present, several noteworthy obstacles hinder the
PET scanning to examine acute histamine-induced itching. effective application of AI in the medical field, especially in
PET scans of AD patients revealed increased brain activity, dermatology. One of the primary challenges is the lack of
especially in the basal ganglia, which are known to be a key
regulator of the itch-scratch cycle. 74 appropriate quality manual annotation and limited sample
numbers in existing dermatology training datasets for AI
A study using arterial spin labeling fMRI, which algorithms. This deficiency diminishes the accuracy and
involved histamine induction in seven healthy controls usefulness of AI algorithms, rendering them inadequate
and eight AD patients, has yielded similar findings, to meet the demands of routine clinical applications.
78
demonstrating markedly enhanced cerebral perfusion Furthermore, AI algorithms, usually developed using pre-
after acute scratching in patients with AD. In addition, existing samples, often fail to align with actual medical
75
two other studies investigating brain imaging in needs, resulting in a disconnect between them and practical
individuals with chronic scratching revealed increased clinical requirements. High-quality training sets are
activity in reward circuits and motor-related brain areas necessary for AI to learn from experience and evolve over
during scratching. 76,77 Moreover, patients with chronic time-a capability that human doctors possess naturally.
itching exhibited reduced sensation of itch in response Without such datasets, AI’s capacity is hampered, making
to scratching in these two studies that used arterial spin it challenging to satisfy the growing expectations of both
labeling fMRI to measure brain activity during active clinical and scientific domains, especially in areas such
scratching. Concurrently, brain regions linked to the as the hairy scalp, mucosal membranes, uncommon skin
central reward system showed significant activation. Taken disorders, and the identification of picture artifacts such
as a whole, these results emphasize the pleasurable aspect as colorful marks and tattoos on the skin. Furthermore,
79
of scratching in relieving both acute and chronic pruritus, training AI to recognize and diagnose a variety of skin
and they suggest the possibility that individuals with AD problems is difficult due to the wide range of dermatological
may develop an addiction to scratching.
diseases and the lack of standard criteria for identification
It is important to recognize three main limitations of and diagnosis. There remains a bottleneck in using AI
80
neurological imaging in the context of AD-related acute for the automatic recognition and diagnosis of various
and chronic pruritus. First, many of these studies involve dermatopathological images, with current AI applications
73
limited sample sizes, and the results are contingent on the more frequently used for differentiating between normal
techniques used to induce itching and perform imaging. and abnormal instances. 81,82 In addition, another major
Second, there is a scarcity of studies demonstrating barrier to the use of AI in dermatology is the presence of
differences in resting states between healthy controls rare disorders, characterized by a low number of cases and
and AD patients. Instead, these techniques need an itch insufficient specimens for appropriate ML training. 83
stimulus other than AD-induced itch (e.g., histamine-
induced itch) to identify changes that might not accurately At the center of these challenges, it lies the critical
mimic the natural state. Third, tests such as PET and fMRI importance of guaranteeing the quality of data used in
are expensive for the health-care system and uncomfortable AI services. This intricacy is compounded by intertwined
for patients, requiring extensive time for execution and issues, such as potential inaccuracies in the accuracy and
analysis. These variables diminish the usefulness of brain reliability of annotations, which can affect the accuracy of
imaging for routine clinical diagnosis and standard care. the model training assumptions. Furthermore, variability
The different distribution of skin lesions with varying introduced in data collection procedures aggravates these
morphologies, intensities, and durations might complicate problems, resulting in datasets that are highly inconsistent
the diagnosis of AD further. However, in recent years, there and pose significant challenges for applying general models
has been a rapid development in AI-based techniques for to a multitude of clinical scenarios. The presence of such
image analysis. One method that represents complicated errors, artifacts, or the lack of proper data preprocessing
patterns is the use of CNNs, which rely on DL algorithms algorithms adds another layer of complexity to raw
to identify correlations between neighboring images and dermatological data, which must be addressed through
integrate them into successive layers. CNNs are used in comprehensive data preprocessing strategies to enhance
image analysis across a variety of medical specialties, the reliability of the model outputs. 84

Volume 1 Issue 2 (2024) 57 doi: 10.36922/aih.2775