Forum for Nordic Dermato-Venereology No 3, 2019 Telemedicine | Page 15

Carsten Sauer Mikkelsen, Kristian Bakke Arvesen, Peter Bjerring and Luit Penninga – Artificial Intelligence in Dermatology
established algorithms, to assist in diagnosis (e.g. the “ABCs”
of melanoma, in addition to texture, patterns, and other
geome­trical features). Fractal analysis has been used for the
diagnosis of other cancers (e.g. mammography for breast
cancer) (19) which has been used by health professionals, and
not by patients for assessment of their own cancer risk, as for
some of the applications identified in the current study (6, 7).
M obile
phone applications
The first group, mobile phone applications, are intended to
provide information about melanoma or non-melanoma
skin cancer, and guidance on whether people should consult
a doctor for a specific lesion that they have photographed
with their smart device. Some applications are also designed
to monitor skin lesions and register whether changes occur
with time (6, 7). The mobile phone and smart device apps can
be divided into 2 groups: those that provide analysis of the
image; and those that produce a store-and-forward image to
be sent to a dermatologist for evaluation (7).
Mobile phone applications using image analysis
One review identified 39 mobile phone apps for melanoma,
18 of which used some image analysis (6). A Cochrane Review
of smartphone apps for triaging adults with skin lesions suspi-
cious for melanoma included 2 studies on image analysis (7).
Both studies had a high risk of bias. Sensitivities in detecting
melanoma ranged from 7% (95% confidence interval (95%
CI) 2–16%) to 73% (95% CI 52–88%) and specificities ranged
from 37% (95% CI 29–46%) to 94% (95% CI 87–97%). For an
application to be safe, it must have almost 100% sensitivity in
order not to miss any diagnoses of melanoma (i.e. no false-neg-
ative cases). Ideally, the application would also have high
specificity (i.e. no false-positive cases), because false-positive
cases cause unnecessary worry among patients. Applications
with very high sensitivity but low specificity will also cause
worry, and lead to many unnecessary referrals and surgery,
thereby increasing the dermatologists’ workload.
Mobile phone applications using store-and-forward of images
Mobile phone apps with store-and-forward of photographic
images were used in 9 out of 39 mobile phone apps on melano-
ma (6). These applications forward a photograph of the lesion
to an experienced professional, such as a dermatologist, for
review, and then communicate a recommendation regarding
the nature of the lesion to the user (6). This might also be
termed a teledermatology service for patients. The Cochrane
Review of smartphone applications for triaging adults with
skin lesions suspicious for melanoma included a study using
this store-and-forward application with high risk of bias (7).
This study had a sensitivity of 98% (95% CI 90–100%) and
Forum for Nord Derm Ven 2019, Vol. 24, No. 3
specificity of 30% (95% CI 22–40%) (7). Thus, one diagnosis
of melanoma was missed (one false-negative case). This ap-
plication had a specificity of only 39%, which would result
in many unnecessary referrals to the dermatologist and many
people worrying unnecessarily and undergoing unnecessary
surgery (7).
Potential advantages in both groups of mobile phone apps are
increased involvement of the public in detecting suspicious
skin changes, including the chance of early detection of
melanoma. Furthermore, increased patient engagement and
involvement may result in better educated patients, and more
effective and efficient consultations (1, 6, 7). Furthermore,
both applications may allow better access to healthcare for
people in remote areas.
However, routine use of these mobile phone apps is not advised
based on current evidence, despite their potential advantages
(7). Mobile phone apps with image analysis are not sufficiently
safe, as they miss cases of melanoma, and, even worse, they
may give patients false reassurance that their lesion is not
cancer, thus delaying diagnosis.
Evidence regarding mobile phone apps with store-and-forward
options to the dermatologist is sparse, and involves only one
study (7). This method seems to be more sensitive; although,
due to low specificity, it might result in many false-positives,
causing unnecessary referrals, surgeries and worries.
Further development, evaluation and research into these appli-
cations is important before they can play a role in healthcare.
It is important to note, however, that these apps are available
to the public, and that patients will use them, even though
they are not safe.
A pplications
to help dermatologists increase the
accuracy of diagnosing malignant melanoma
The second group is applications that help dermatologists to
increase the accuracy of diagnosing malignant melanoma.
Diagnosis of malignant melanoma is very important, but
also very difficult (8). Visual diagnosis is not easy, is highly
observer-dependent, and may be associated with low accuracy
when performed by young and unexperienced dermatologists.
The accuracy of experts is between 75% and 84% (5, 8). Hence,
other tests that may facilitate the diagnosis of melanoma in a
specialist setting have been developed (5, 8–16, 19–23). These
include reflectance confocal microscopy, optical coherence
tomography, high frequency ultrasound, as well as comput-
er-assisted diagnosis or AI-based applications (8–16, 19–23).
Histological confirmation of malignant melanoma remains
the gold standard (8). AI-based applications may provide a
T heme I ssue : T eledermatology
99