Original ResearchMonitoring patients at risk for melanoma: May convolutional neural networks replace the strategy of sequential digital dermoscopy?
Introduction
Melanoma still accounts for most skin cancer deaths, and incidence remains on a high level [1]. On the one hand, early diagnosis is of utmost importance because surgical excision is curative in thin melanomas [2]. On the other hand, overdiagnosis of melanoma due to increased diagnostic scrutiny, lowered thresholds for surgical biopsies, and lowered thresholds to diagnose melanoma by dermatopathologists is a serious concern [3]. Therefore, new technological advances should not only solely focus on ‘not missing melanoma’ but also on ‘reducing unnecessary excisions’. Compared with the unaided eye, dermoscopy improves the diagnostic accuracy for melanoma particularly when applied by experienced dermatologists [4]. Dermoscopic patterns of melanoma and various diagnostic algorithms have been defined, yet early or featureless melanomas remain difficult to diagnose [5,6]. Because changes in lesion pattern and size frequently indicate malignancy, it is reasonable to monitor pigmented lesions over time. Sequential digital dermoscopy (SDD) allows documentation of dynamic changes and facilitates melanoma detection, while avoiding unnecessary excisions of nevi [7,8]. Monitoring patients by SDD was shown to be particularly effective in patients at higher risk for melanoma, e.g. in patients affected by atypical mole syndrome or familial atypical multiple mole melanoma syndrome [9,10]. Short-term SDD follow-up is helpful for making a correct management decision in a suspicious lesion lacking clear melanoma features, whereas long-term follow-up focuses on monitoring of high-risk patients and also includes inconspicuous lesions [7,8,11]. We previously found that a significant proportion of melanomas diagnosed in the course of SDD could not be identified by dermatologists in baseline images, which emphasises the diagnostic value added by using SDD in high-risk patients [12].
More recently, there have been numerous reports on the high-level diagnostic accuracy of convolutional neural networks (CNN) for skin cancer detection in clinical and dermoscopic images [[13], [14], [15], [16], [17]]. With the present study, we aimed to investigate whether current CNN may replace the strategy of SDD.
Section snippets
Images for evaluation
Our study was approved by the local ethics committee and performed in accordance with the Declaration of Helsinki principles. Digital dermoscopy images of 59 patients at increased risk for melanoma were selected from a database. All patients were monitored by SDD. For every patient, a baseline image quartet of four dermoscopic images was arranged, including three benign nevi and one lesion identified as melanoma during SDD. Moreover, follow-up quartets were created, including a follow-up image
Characteristics of imaged lesions and dermatologists
For each of the 59 patients monitored by SDD, we created a baseline and follow-up image quartet. Fig. 1 depicts representative quartets. At the time of excision, 28 (47.5%) melanomas were in situ, and 31 (52.5%) melanomas were invasive with a mean thickness of 0.4 mm (range: 0.2–1.3 mm). Sixteen (27.1%) melanomas had developed in association with pre-existing nevi. Of the 59 melanomas, 32 (54.2%) melanomas were located on the trunk, 24 (40.7%) melanomas on the extremities and three melanomas
Discussion
Previous studies have clearly shown that deep learning CNN are capable of classifying skin lesions on or above the level of trained dermatologists [[13], [14], [15]]. However, exploring the limitations and benefits of CNN when applied in different clinical scenarios is still an ongoing process [[21], [22], [23]]. With the study, we investigated whether monitoring melanocytic lesions by SDD might be replaced by CNN assessment of baseline lesions.
SDD was shown to be effective for early melanoma
Ethics
Reviewed and approved by the ethic committee of the medical faculty of the University of Heidelberg (Approval number S-629/2017).
Funding
No funding. This research received no specific grant from any public, commercial or not-for-profit sector.
Authors’ contributions
Study concepts: Haenssle, Kittler, Tschandl, Study design: Haenssle, Kittler, Tschandl, Data acquisition: Haenssle, Kittler, Tschandl, Quality control of data and algorithms: Haenssle, Kittler, Tschandl, Winkler, Data analysis and interpretation: Haenssle, Kittler, Tschandl, Winkler, Statistical analysis: Haenssle, Tschandl, Winkler, Manuscript preparation: Haenssle, Tschandl, Winkler, Manuscripts editing: Haenssle, Tschandl, Winkler, Manuscript review: Enk, Fink, Haenssle, Kittler, Sies,
Access to data and data analysis
HA Haenssle and JK Winkler had full access to all the data in the study and take responsibility for the integrity of the data and the accuracy of the data analysis. Fotofinder Systems was not involved in the design of the study or interpretation of the acquired data.
Study registration
This study is registered at the German Clinical Trial Register (DRKS): Study-ID DRKS00013570.
Financial Disclosure
None.
Conflict of interest statement
P Tschandl reports grants from MetaOptima and Lilly, consulting fees from Silverchair, and speaker honoraria from Lilly and FotoFinder, outside the submitted work. A Enk reports a DFG grant, honoraria from Biotest, MSD, Janssen-Cilag and BMS, participation in an MSD Data Safety Monitoring Board and leadership in the EDF and IDSI outside the submitted work. H Kittler states honoraria from Fotofinder and Pelpharma, as well as receipt of equipment from Fotofinder, Heine, 3Gen, Derma Medical
Acknowledgement
The authors would like to acknowledge Mohamed Souhayel Abassi and Tobias Fuchs (Department of Research and Development, FotoFinder Systems GmbH, Bad Birnbach, Germany) for their technical support regarding CNN assessment.
References (29)
- et al.
Melanoma
Lancet
(2018) - et al.
European consensus-based interdisciplinary guideline for melanoma. Part 2: treatment–update 2019
Eur J Cancer
(2020) - et al.
Dermoscopy of pigmented skin lesions
J Am Acad Dermatol
(2005) - et al.
Comparison of the accuracy of human readers versus machine-learning algorithms for pigmented skin lesion classification: an open, web-based, international, diagnostic study
Lancet Oncol
(2019) - 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) - et al.
Man against machine reloaded: performance of a market-approved convolutional neural network in classifying a broad spectrum of skin lesions in comparison with 96 dermatologists working under less artificial conditions
Ann Oncol
(2020) - et al.
Clinical outcomes from skin screening clinics within a community-based melanoma screening program
J Am Acad Dermatol
(2006) - et al.
Melanoma recognition by a deep learning convolutional neural network-Performance in different melanoma subtypes and localisations
Eur J Cancer
(2020) - et al.
Association between different scale bars in dermoscopic images and diagnostic performance of a market-approved deep learning convolutional neural network for melanoma recognition
Eur J Cancer
(2021) - et al.
Robustness of convolutional neural networks in recognition of pigmented skin lesions
Eur J Cancer
(2021)
The rapid rise in cutaneous melanoma diagnoses
N Engl J Med
Dermoscopy compared with naked eye examination for the diagnosis of primary melanoma: a meta-analysis of studies performed in a clinical setting
Br J Dermatol
Limitations of dermoscopy in the recognition of melanoma
Arch Dermatol
Identification of clinically featureless incipient melanoma using sequential dermoscopy imaging
Arch Dermatol
Cited by (6)
Noninvasive imaging techniques in dermato-oncology
2023, Best Practice OnkologieRecent Advances in Melanoma Diagnosis and Prognosis Using Machine Learning Methods
2023, Current Oncology ReportsArtificial Intelligence in Dermatology: Challenges and Perspectives
2022, Dermatology and Therapy