Introduction: Dermatofluoroscopy is an optical noninvasive method of melanoma/nevus differentiation that has shown 89% sensitivity and 45% specificity in clinical trials, but long measurement duration hinders clinical use. An intelligent algorithm was developed to shorten the measurement time without compromising its diagnostic accuracy. It uses dermoscopic images of the skin lesions to be measured to select measurement points based on the assessment of color values. Methods: 27 patients with a total of 29 lesions suggestive of cutaneous melanoma were included in a clinical study and measured with both methods, conventional dermatofluoroscopy and the newly developed intelligent algorithm. The results were compared to the independent findings of two histopathologists to evaluate diagnostic accuracy and time saved. Results: There was a median reduction of measurement points from 265 to 158 (40%). Meanwhile, the intelligent algorithm showed a higher diagnostic accuracy than conventional dermatofluoroscopy (area under the curve of 72% vs. 63%). Conclusion: The intelligent algorithm did not perform inferior to the conventional method while saving 40% of time. However, measurement times remain long compared to other noninvasive methods of diagnosing malignant melanoma. Further studies are needed to evaluate clinical suitability.

The societal importance of cutaneous malignant melanoma (MM) is highlighted by its epidemiology: In 2022, there were an estimated 100,000 new cases and 7,500 deaths in the USA [1]. While the disease can be curatively excised in its early stages, late stages can be fatal and require systemic treatment [2]. One cornerstone in reducing melanoma morbidity and mortality, therefore, is early diagnosis. The gold standard for diagnosis is total excision followed by histopathological assessment [3]. However, excising all pigmented skin lesions (PSLs) in every patient is neither feasible nor desirable, considering that the average patient presents with 13 PSL and some patients have considerably more (calculated with reference [4]). Therefore, as many PSLs as possible should ideally be ruled out from being MM by noninvasive methods. The preferred method of non-invasive melanoma diagnosis is dermatoscopy, which is a handheld reflected light microscope, allowing for a more detailed examination of skin lesions. While estimations of sensitivity and specificity vary between 50% and 90%, it is agreed that the level of examiner experience plays a decisive role [5‒7]. Due to the subjectivity of dermatoscopy and its requirement for expertise, there have been attempts to establish objective methods of noninvasive melanoma diagnosis, such as dermatofluoroscopy (DF) [8]. In DF, 800 nm, 1 ns laser pulses are used to specifically excite melanin fluorescence, while endogenous fluorophores with otherwise bigger quantum yield are minimally affected [9]. The resulting fluorescence spectrum indicates whether a lesion is benign or malignant, with malignant spectra being red-shifted [9]. A support vector machine classifies each resulting spectrum [10]. For overall lesion classification, a DF score is calculated by multiplying the proportion of malignant spectra among all measured spectra with a factor of 37.5 and compared to a cut-off value. This cut-off can be adjusted to balance sensitivity and specificity, with literature suggesting a cut-off value of 28, corresponding to about 75% malignant spectra [8, 10, 11]. A clinical drawback of DF is the duration of the scan defined as: (duration of one scan point + duration of movement of the motors between scans) × number of scans, which amounts to approximately 21 min for a circular 6 mm lesion. Reducing the number of points to be measured can decrease the acquisition time. Retrospective analysis of DF measurements from previous studies has shown a relationship between the colors in pre-scan dermoscopic images and the likelihood of identifying spectra characteristic of either MM or nonmelanoma (NM). Certain colors demonstrate better class separability, meaning the DF scores of MM and NM are on average more distinct when measured on these colors. Conversely, measurement points on colors with poor class separability tend to yield similar DF scores regardless of whether the lesion is MM or NM, thus offering limited diagnostic value. Consequently, an algorithm was developed that uses pre-scan dermoscopic images, taken immediately before the DF measurement, to select fewer but more informative measurement points. This clinical study aims to determine whether the new algorithm successfully reduces measurement time and, secondarily, how it affects sensitivity and specificity.

Study Design and Patients

For this clinical study, 27 patients with a total of 29 PSLs, clinically indicated to be a MM and planned for excision, were recruited. The study was conducted at the Charité – Universitätsmedizin Berlin according to the Declaration of Helsinki as revised in 2013. Approval was obtained by the Local Ethics Committee (EA1/267/21), and the study was registered in the German Clinical Trials Register (DRKS00028051). All subjects were of legal age and gave written informed consent.

Investigational Device and Diagnostic Algorithm

Measurements were performed using a DermaFC CE-certified medical device of class IIa, provided by Magnosco (Magnosco GmbH, Berlin, Germany). It consists of a roller-mounted main body with a touch screen monitor and movable arm with a scanning head. To examine a PSL, the scanning head is attached to the skin with an adhesive ring. The scanned area is shielded from ambient light by a cover. Both conventional DF [8] and the new algorithm use a pre-scan photo, i.e., dermoscopic-quality image of the lesion, to allow selection of the lesion outlines via the touch monitor. Conventional DF then scans the lesion on a regular grid of 200-µm step-size, resulting in hundreds of measurement points. In contrast, the new algorithm specifically selects certain measurement points based on their color in the pre-scan photo. To obtain the fluorescence spectra, the skin lesion to be diagnosed was scanned with a laser focused in the dermal-epidermal junction (wavelength 800 nm, pulse duration 1 ns, repetition rate 1.2 kHz, intensity ≈0.35 GW/cm2). Spectrally and spatially resolved signals from photoexcited skin fluorophores were detected with a cooled CCD camera (in the 380–780 nm range) and analyzed with a support vector machine model. The measured signals of molecules, such as melanin, NADH, FAD or keratin, reflect the cancer-induced alterations in the microenvironment of the skin tumor and therefore, can be used for discriminating malignant PSL from their benign simulators.

Data Acquisition

All lesions were measured using both, conventional DF and the new intelligent algorithm. All included patients were shaved at the PSL site. Before DF measurement began, window blinds were closed, and light was switched off to minimize light interference. The dermatofluoroscope took a pre-scan dermoscopic image that allows measurement area selection. To accelerate data collection, the measurement points of both, the conventional and the new algorithm, were placed on a shared grid to perform only one scan. Thus, exactly the same area was measured in both methods. This approach also saved time as overlapping points were only measured once.

Results from both scanning approaches were compared. For the histopathological diagnosis, each of the excised PSL was examined by two independent, blinded pathologists. In case of discordance, a third independent pathologist was consulted.

Selection of Measurement Points and Color Clusters

Prior to measurement, the dermoscopic pre-scan photo is used to generate an optimized grid which differs from the conventional grid by increasing the measurement point density on colors with a high-class separability and decreasing it on colors with a low class separability. The measurement points’ degree of separability was assessed by matching DF scores of measurement points on previously diagnosed lesions to the colors on their pre-scan photos. The colors were then grouped into color clusters (online suppl. Fig. 1; for all online suppl. material, see https://doi.org/10.1159/000542854) [12] and expected DF scores were calculated for each cluster. For every color cluster, an expected DF score was calculated for MM and NM, respectively. The better the separability between the expected score of MM and NM, the more definitive the differentiation between MM and its benign simulator (e.g., dysplastic nevus) and the more useful the measurement point. An expected overall score for MM and NM is calculated for every lesion based on the number of measurement points per color cluster. An individual (i.e., lesion-specific) cut-off value is then calculated allowing for optimal separability. Due to the small sample size, the categorization into color clusters was performed with a leave-one-out approach, i.e., for every single lesion analyzed, the colors and DF scores of all other lesions were used. The colors on every lesion were matched to the color clusters by an algorithm that was composed of the data points on all other lesions.

Statistical Analysis

Descriptive statistics and frequency tabulations were used to summarize the demographic and clinical characteristics of patients and their PSLs. The conventional method, the intelligent algorithm, and its different configurations were analyzed using sensitivity, specificity, and the area under the receiver operating characteristic curve. The reduction of measurement points was compared with the Wilcoxon signed-rank test, and results with p < 0.05 were considered significant.

Optimization of the Grid

While conventional DF uses a regular grid with 200-µm intervals, the intelligent algorithm uses an irregular grid based on the class separability of measurement points. This grid can be optimized for different goals: selecting more points for higher diagnostic accuracy or fewer points for time efficiency. The trade-off between diagnostic accuracy and time savings is managed by a parameter g, which balances these two objectives during the grid’s iterative optimization. The following formula shows how the factor g influenced the optimization procedure. Various values of g were evaluated, along with different numbers of color clusters nCl.
where:
  • N_new(color_i): the new number of measurement points after the optimization step for a given color cluster i.

  • Δ separability: the change in class separability of the lesion.

  • g: the factor weighting Δ separability against scan duration.

  • N_all: the total number of points to be measured.

  • lr: the learning rate.

  • N(color_i): the number of points in a color cluster i.

The formula above represents a step in an iterative optimization procedure, which is initialized from points distributed according to the regular 200-µm grid. For each color cluster, the function calculates the number of points to be added or removed, with a learning rate that allows for the adjustment of up to 15 measurement points per iteration per cluster. This process, known as an optimization step, is repeated until no further points are added or removed, or a maximum of 75 optimization steps is reached.

Descriptive Results

Twenty-seven patients (13 female, 14 male) were recruited. 4 out of 27 (15%) had a previous cutaneous melanoma. None presented with a known family history of melanoma. Median lesion diameter was 9 mm (Q1 5 mm, Q3 18 mm). Fitzpatrick skin type varied between type I and III, with 24 (89%) showing type II. Mean age was 64 ± 15 years. 14 patients (52%) received a subsequent histological diagnosis of cutaneous melanoma. These patients had a mean age of 70 ± 13 years and a median lesion diameter of 15 mm (Q1: 7 mm, Q3: 22 mm).

Diagnostic Accuracy of Conventional Grid

For the conventional 200-µm grid, the optimal operating point was at a cut-off of 14, meaning all lesions with a score >14 were considered melanoma. This leads to a sensitivity of 93%, 13 of 14 melanoma were correctly identified with the conventional grid. 6 of 15 NM were identified as such, yielding a specificity of 40%.

Diagnostic Accuracy of Intelligent Algorithm

Two parameters of the algorithm were adjusted for optimal configuration. nCl is the number of color clusters that is used to analyze the scanned lesion. For the current data, three different values for nCl were analyzed: 7, 9 and 12. nCl = 12 showed the lowest diagnostic accuracy with 62%, here operationalized as area under the ROC curve, with a 33% time decrease. nCl = 7 led to a diagnostic accuracy of 68% with a small time decrease of 16%. nCl = 9 had a diagnostic accuracy of 72% while still saving 27% time and was therefore chosen as the preferred nCl.

g determines the trade-off between time and diagnostic accuracy, i.e., number of points saved and area under the corresponding ROC curve. g = 0.001 shows the best balance between the two. While g = 0.0001 shows the highest area under the curve (AUC) of 76%, it only saves 1% time. g = 0.001 has the second highest AUC of 72% with a 27% time decrease (Fig. 1). The practical use of higher values of g is limited as too low numbers of spectra would lead to an unpredictable increase in the uncertainty of diagnostic accuracy. Therefore, the values of g >0.001 were used only to calculate a potential decrease of measurement time. For this study, the preferred configuration of the intelligent algorithm was determined to be a nCl = 9 and g = 0.001, with the most promising operating point at a sensitivity of 100% (14/14) and a specificity of 40% (6/15).

Fig. 1.

Receiver operating characteristic comparing the conventional 200 µm grid (black) and different configurations of the intelligent algorithm. Left shows different numbers of color clusters (nCl) at fixed g = 0.001. Right shows different values of g at a fixed number of color clusters nCl = 9. The AUC and the percentage of measurement points saved compared to conventional DF are shown in the bottom right boxes. AUC, area under the curve.

Fig. 1.

Receiver operating characteristic comparing the conventional 200 µm grid (black) and different configurations of the intelligent algorithm. Left shows different numbers of color clusters (nCl) at fixed g = 0.001. Right shows different values of g at a fixed number of color clusters nCl = 9. The AUC and the percentage of measurement points saved compared to conventional DF are shown in the bottom right boxes. AUC, area under the curve.

Close modal

Comparison of Conventional DF and Intelligent Algorithm

To compare the diagnostic accuracy of the conventional method and the intelligent algorithm, the AUC can be used since it reflects sensitivity as well as specificity at every possible operating point. The AUC of the conventional method was at 0.63, while the intelligent algorithm at nCl = 9 and g = 0.001 had an AUC of 0.72, therefore, not performing inferiorly to the conventional method.

Online supplementary Table 1 provides an overview of all lesions, their histopathological diagnosis, as well as the scores from the conventional and intelligent prediction at the configuration of nCl = 9 and g = 0.001. For the conventional method, the cut-off was set at 14, while for the intelligent algorithm individual cut-offs ranged from 4 to 26. Lesion 14 (Lentigo maligna) was falsely classified as NM by the conventional method, but correctly classified as MM by the intelligent method. The intelligent algorithm did not miss any melanoma. The intelligent algorithm improved the sensitivity while the specificity remained unchanged (two false positives were removed but two new false positives were added, so that both methods falsely classified 9 NM as MM).

To evaluate the potential time to be saved by the intelligent algorithm, the number of measurement points used by the intelligent algorithm was compared to the number of points used by the conventional method (online suppl. Table 1; Fig. 2) [12]. For 7 of 29 lesions (24%), the intelligent algorithm at nCl = 9 and g = 0.001 required more measurement points than the conventional method. For the remaining 22 lesions, it required less. Q1, Q2, and Q3 were 217, 265, and 344 measurement points per lesion for the conventional method and 121, 158, and 196 for the intelligent algorithm, respectively. This results in a median decrease of measurement points of 40% (158/265). The median scanning time decreases from 8 min to 50 s to 6 min and 16 s. The Wilcoxon signed-rank test resulted in p = 0.0015. online supplementary Figure 2 [12] highlights the decrease in measurement points graphically. Figure 2 compares the distribution of measurement points of the conventional method and the intelligent algorithm. They show how the intelligent algorithm increases the measurement points in relevant areas while less relevant areas of the lesion are measured more sparsely.

Fig. 2.

Comparison of the distribution of measurement points for lesion 25 (above), a superficial spreading melanoma, and lesion 16 (below), a melanocytic nevus. Left shows the pre-scan photo, middle shows the measurement points in the conventional 200 µm grid and right shows measurement points as placed by the intelligent algorithm (in the nCl = 9 and g = 0.001 configuration). Note how healthy skin and lighter brown tones are almost fully ignored in favor of a higher density of measurement points on colors that are more useful for class separability.

Fig. 2.

Comparison of the distribution of measurement points for lesion 25 (above), a superficial spreading melanoma, and lesion 16 (below), a melanocytic nevus. Left shows the pre-scan photo, middle shows the measurement points in the conventional 200 µm grid and right shows measurement points as placed by the intelligent algorithm (in the nCl = 9 and g = 0.001 configuration). Note how healthy skin and lighter brown tones are almost fully ignored in favor of a higher density of measurement points on colors that are more useful for class separability.

Close modal

This clinical study introduces and evaluates an intelligent algorithm for DF that aims to reduce the measurement points and thereby, scanning time. Different configurations of the intelligent algorithm were compared. The best configuration was then compared to the conventional grid method. The goal was a significant point reduction without sacrificing diagnostic accuracy, i.e., AUC. The intelligent algorithm did not show inferior diagnostic accuracy compared to the conventional method with respective AUC of 0.72 and 0.63. The intelligent algorithm showed a median decrease of 40% in measurement points compared to the conventional algorithm.

Diagnostic Accuracy of Conventional 200 µm Grid

In the biggest previous study, DF with the conventional grid showed a sensitivity of 89% and a specificity of 45% [8]. The current study showed a similar performance, with 93% sensitivity and 40% specificity.

Diagnostic Accuracy of Intelligent Algorithm

Color clusters were calculated with the data of this study using a leave-one-out approach. The results are therefore not prospective. Based on diagnostic accuracy, i.e., AUC, and measurement points saved, nCl = 9 was then selected as the best number of color clusters. It remains unclear if nCl = 9 would also be ideal for future studies with the intelligent algorithm. Hypothetically, larger data sets would allow for higher nCl without risk of overfitting the model.

g = 0.001 was chosen as the optimal compromise between diagnostic accuracy and reduction of measurement points. It is unclear how this configuration of g would translate to future studies.

Comparison with Other Noninvasive Methods of Melanoma Diagnostics

Given the clinical importance of early and objective melanoma detection, numerous noninvasive methods have been investigated, including Raman spectroscopy [13], confocal laser microscopy [14], electrical impedance spectroscopy [15], optical coherence tomography [16], and multiphoton tomography [17]. The most attention has been directed toward photo-based AI methods, which evaluate lesions based on their dermoscopic appearance. In a recent study, a dermatoscopy-based algorithm achieved a sensitivity of 97% and a specificity of 37% [18]. These results are similar to the respective 100% and 40% achieved by the intelligent algorithm in this study; however, the comparability is limited due to post-study fine-tuning of the color model. A meta-analysis reported a sensitivity of 91% and a specificity of 79% for dermatoscopy-based AI algorithms [19]. Another meta-analysis reported a sensitivity of 74% and a specificity of 84%, with these numbers decreasing to 51% and 83% when independent training and test sets were used [20].

The deep learning-based classification algorithms employed in dermatoscopy currently lack explainability [18, 21], meaning the rationale behind a lesion being classified as benign or malignant remains unclear. In contrast, classification by DF is a linear function of the proportion of malignant spectra, with the intelligent algorithm only preselecting measurement points based on their color. Moreover, the method’s acceptance among patients is high; in a recent study, 74% of respondents considered DF trustworthy [22], and this might be further increased with faster measurement times.

However, DF remains slower compared to competing methods like AI-based dermatoscopy and electrical impedance spectroscopy, which generate results within seconds [23]. Additionally, DF is limited to diagnosing PSLs, whereas non-melanin-dependent methods can also be used for basal cell carcinoma and other non-pigmented skin tumors [24, 25]. Other drawbacks of DF include its sensitivity to patient movement and light interference, as well as the device being larger and less portable compared to dermatoscopy or electrical impedance spectroscopy.

The new intelligent algorithm evaluated in this study maintained a high diagnostic accuracy while saving 40% of measurement points per lesion, directly translatable into saved time. Thus, it outperformed the conventional grid scanning method. Further research is required to evaluate the prospective performance of the intelligent scanning algorithm. Widespread clinical application seems unlikely at this stage as DF, though faster with the intelligent algorithm, is still significantly slower than competing methods without significantly outperforming them in terms of diagnostic accuracy.

This study protocol was reviewed and approved by the Ethics Committee of the Charité – Universitätsmedizin Berlin, Approval No. EA1/267/21 on 29th of September 2021. All subjects were of legal age and gave written informed consent. The study is registered with the German Clinical Trials Register with the ID DRKS00028051.

Lukasz Szyc and Eduard Dronnik are employees of Magnosco, a private company that develops the method of DF.

Magnosco GmbH provided funding for this study.

The measurements and data acquisition were performed by Karl Weihmann. Johannes Schleusener, Thomas Eigentler, Lukasz Szyc, and Martina Meinke were responsible for the conception of the study. The design was finalized by Johannes Schleusener, Franziska Ghoreschi, Rose Moritz, Lukasz Szyc, Eduard Dronnik, and Martina Meinke. Histopathologic analysis was performed by Franziska Ghoreschi and Rose Moritz. Results were analyzed and evaluated by Karl Weihmann, Lukasz Szyc, and Eduard Dronnik. The article draft was written by Karl Weihmann. All authors contributed to reviewing. All authors give their final approval and agree to be accountable for all aspects of the work.

All data generated or analyzed during this study are included in this article. Further inquiries can be directed at the corresponding author.

1.
Institute NC
.
Cancer stat facts: melanoma of the skin
.
2022
.
2.
Skudalski
L
,
Waldman
R
,
Kerr
PE
,
Grant-Kels
JM
.
Melanoma: an update on systemic therapies
.
J Am Acad Dermatol
.
2022
;
86
(
3
):
515
24
.
3.
Swetter
SM
,
Tsao
H
,
Bichakjian
CK
,
Curiel-Lewandrowski
C
,
Elder
DE
,
Gershenwald
JE
, et al
.
Guidelines of care for the management of primary cutaneous melanoma
.
J Am Acad Dermatol
.
2019
;
80
(
1
):
208
50
.
4.
Mackie
RM
,
English
J
,
Aitchison
TC
,
Fitzsimons
CP
,
Wilson
P
.
The number and distribution of benign pigmented moles (melanocytic naevi) in a healthy British population
.
Br J Dermatol
.
1985
;
113
(
2
):
167
74
.
5.
Ignazio
S
,
Aldo
B
,
Luigi
C
,
Roberto
G
,
Alberto
L
,
Serena
M
, et al
.
Computer-aided diagnosis of melanocytic lesions
.
Anticancer Res
.
2005
;
25
(
6C
):
4577
.
6.
Haenssle
HA
,
Fink
C
,
Schneiderbauer
R
,
Toberer
F
,
Buhl
T
,
Blum
A
, et al
.
Man against machine: diagnostic performance of a deep learning convolutional neural network for dermoscopic melanoma recognition in comparison to 58 dermatologists
.
Ann Oncol
.
2018
;
29
(
8
):
1836
42
.
7.
Robsahm
TE
,
Helsing
P
,
Svendsen
HL
,
Veierød
MB
.
Clinical suspicion sensitivity of nodular and superficial spreading melanoma
.
Acta Derm Venereol
.
2021
;
101
(
4
):
adv00427
.
8.
Forschner
A
,
Keim
U
,
Hofmann
M
,
Spankuch
I
,
Lomberg
D
,
Weide
B
, et al
.
Diagnostic accuracy of dermatofluoroscopy in cutaneous melanoma detection: results of a prospective multicentre clinical study in 476 pigmented lesions
.
Br J Dermatol
.
2018
;
179
(
2
):
478
85
.
9.
Eichhorn
R
,
Wessler
G
,
Scholz
M
,
Leupold
D
,
Stankovic
G
,
Buder
S
, et al
.
Early diagnosis of melanotic melanoma based on laser-induced melanin fluorescence
.
J Biomed Opt
.
2009
;
14
(
3
):
034033
.
10.
Szyc
L
,
Hillen
U
,
Scharlach
C
,
Kauer
F
,
Garbe
C
.
Diagnostic performance of a support vector machine for dermatofluoroscopic melanoma recognition: the results of the retrospective clinical study on 214 pigmented skin lesions
.
Diagnostics
.
2019
;
9
(
3
):
103
.
11.
Szyc
L
,
Scharlach
C
,
Haenssle
H
,
Fink
C
.
In vivo two-photon-excited cellular fluorescence of melanin, NAD(P)H, and keratin enables an accurate differential diagnosis of seborrheic keratosis and pigmented cutaneous melanoma
.
J Biomed Opt
.
2021
;
26
(
7
):
075002
.
12.
Weihmann
K
.
In vivo study to evaluate an intelligent algorithm for time efficient detection of malignant melanoma using dermatofluoroscopy
.
Mendeley Data, V1
;
2023
.
13.
Schleusener
J
,
Gluszczynska
P
,
Reble
C
,
Gersonde
I
,
Helfmann
J
,
Fluhr
JW
, et al
.
In vivo study for the discrimination of cancerous and normal skin using fibre probe-based Raman spectroscopy
.
Exp Dermatol
.
2015
;
24
(
10
):
767
72
.
14.
Alarcon
I
,
Carrera
C
,
Palou
J
,
Alos
L
,
Malvehy
J
,
Puig
S
.
Impact of in vivo reflectance confocal microscopy on the number needed to treat melanoma in doubtful lesions
.
Br J Dermatol
.
2014
;
170
(
4
):
802
8
.
15.
Braun
RP
,
Mangana
J
,
Goldinger
S
,
French
L
,
Dummer
R
,
Marghoob
AA
.
Electrical impedance spectroscopy in skin cancer diagnosis
.
Dermatol Clin
.
2017
;
35
(
4
):
489
93
.
16.
Gambichler
T
,
Schmid-Wendtner
MH
,
Plura
I
,
Kampilafkos
P
,
Stücker
M
,
Berking
C
, et al
.
A multicentre pilot study investigating high-definition optical coherence tomography in the differentiation of cutaneous melanoma and melanocytic naevi
.
J Eur Acad Dermatol Venereol
.
2015
;
29
(
3
):
537
41
.
17.
Seidenari
S
,
Arginelli
F
,
Dunsby
C
,
French
PMW
,
König
K
,
Magnoni
C
, et al
.
Multiphoton laser tomography and fluorescence lifetime imaging of melanoma: morphologic features and quantitative data for sensitive and specific non-invasive diagnostics
.
PLoS One
.
2013
;
8
(
7
):
e70682
.
18.
Marchetti
MA
,
Cowen
EA
,
Kurtansky
NR
,
Weber
J
,
Dauscher
M
,
DeFazio
J
, et al
.
Prospective validation of dermoscopy-based open-source artificial intelligence for melanoma diagnosis (PROVE-AI study)
.
NPJ Digit Med
.
2023
;
6
(
1
):
127
.
19.
Rajpara
SM
,
Botello
AP
,
Townend
J
,
Ormerod
AD
.
Systematic review of dermoscopy and digital dermoscopy/artificial intelligence for the diagnosis of melanoma
.
Br J Dermatol
.
2009
;
161
(
3
):
591
604
.
20.
Dick
V
,
Sinz
C
,
Mittlböck
M
,
Kittler
H
,
Tschandl
P
.
Accuracy of computer-aided diagnosis of melanoma: a meta-analysis
.
JAMA Dermatol
.
2019
;
155
(
11
):
1291
9
.
21.
Ibrahim
R
,
Shafiq
MO
.
Explainable convolutional neural networks: a taxonomy, review, and future directions
.
ACM Comput Surv
.
2023
;
55
(
10
):
1
37
.
22.
Fink
C
,
Uhlmann
L
,
Hofmann
M
,
Forschner
A
,
Eigentler
T
,
Garbe
C
, et al
.
Patientenakzeptanz und -vertrauen in die automatisierte, computergestützte Diagnostik des Melanoms mithilfe der Dermatofluoroskopie
.
J Dtsch Dermatol Ges
.
2018
;
16
(
7
):
854
60
.
23.
Malvehy
J
,
Hauschild
A
,
Curiel-Lewandrowski
C
,
Mohr
P
,
Hofmann-Wellenhof
R
,
Motley
R
, et al
.
Clinical performance of the Nevisense system in cutaneous melanoma detection: an international, multicentre, prospective and blinded clinical trial on efficacy and safety
.
Br J Dermatol
.
2014
;
171
(
5
):
1099
107
.
24.
Dua
R
,
Beetner
DG
,
Stoecker
WV
,
Wunsch
DC
.
Detection of basal cell carcinoma using electrical impedance and neural networks
.
IEEE Trans Biomed Eng
.
2004
;
51
(
1
):
66
71
.
25.
Widaatalla
Y
,
Wolswijk
T
,
Adan
F
,
Hillen
LM
,
Woodruff
HC
,
Halilaj
I
, et al
.
The application of artificial intelligence in the detection of basal cell carcinoma: a systematic review
.
J Eur Acad Dermatol Venereol
.
2023
;
37
(
6
):
1160
7
.