If you don't remember your password, you can reset it by entering your email address and clicking the Reset Password button. You will then receive an email that contains a secure link for resetting your password
If the address matches a valid account an email will be sent to __email__ with instructions for resetting your password
Skin Cancer Unit, German Cancer Research Center (DKFZ), Heidelberg, GermanyDepartment of Dermatology, Venereology and Allergology, University Medical Center Mannheim, Ruprecht-Karl University of Heidelberg, Mannheim, Germany
Corresponding author: National Center for Tumor Diseases (NCT), German Cancer Research Center (DKFZ), Im Neuenheimer Feld 460, 69120 Heidelberg, Germany. Fax: +496221 566967.
Department of Dermatology, University Hospital Heidelberg, Heidelberg, GermanyNational Center for Tumor Diseases (NCT), German Cancer Research Center (DKFZ), Im Neuenheimer Feld 460, 69120 Heidelberg, Germany
Thirty percent to 50% of all melanomas and more than half of those in young patients evolve from initially benign lesions.
•
Neither clinicians nor artificial intelligence (AI) algorithms are yet able to reliably predict a nevus' oncologic transformation.
•
AI trained on prospective image data could predict melanoma before it occurs and enhance the accuracy of diagnoses.
Abstract
Recent research revealed the superiority of artificial intelligence over dermatologists to diagnose melanoma from images. However, 30–50% of all melanomas and more than half of those in young patients evolve from initially benign lesions. Despite its high relevance for melanoma screening, neither clinicians nor computers are yet able to reliably predict a nevus’ oncologic transformation. The cause of this lies in the static nature of lesion presentation in the current standard of care, both for clinicians and algorithms. The status quo makes it difficult to train algorithms (and clinicians) to precisely assess the likelihood of a benign skin lesion to transform into melanoma. In addition, it inhibits the precision of current algorithms since ‘evolution’ image features may not be part of their decision. The current literature reveals certain types of melanocytic nevi (i.e. ‘spitzoid’ or ‘dysplastic’ nevi) and criteria (i.e. visible vasculature) that, in general, appear to have a higher chance to transform into melanoma. However, owing to the cumulative nature of oncogenic mutations in melanoma, a more fine-grained early morphologic footprint is likely to be detectable by an algorithm. In this perspective article, the concept of melanoma prediction is further explored by the discussion of the evolution of melanoma, the concept for training of such a nevi classifier and the implications of early melanoma prediction for clinical practice. In conclusion, the authors believe that artificial intelligence trained on prospective image data could be transformative for skin cancer diagnostics by (a) predicting melanoma before it occurs (i.e. pre-in situ) and (b) further enhancing the accuracy of current melanoma classifiers. Necessary prospective images for this research are obtained via free mole-monitoring mobile apps.
Skin cancer is the most common cancer in white-skinned populations. Malignant melanoma, a skin cancer deriving from melanocytes, is accountable for most skin cancer–related deaths [
Early detection greatly improves the prognosis of patients with melanoma; however, despite the use of dermoscopy, dermatologists only rarely achieve sensitivities greater than 80% [
] were the first to report a deep-learning convolutional neural network (CNN) image classifier that performed on par with 21 board-certified dermatologists when identifying images with malignant lesions [
A convolutional neural network trained with dermoscopic images performed on par with 145 dermatologists in a clinical melanoma image classification task.
Comparison of the accuracy of human readers versus machine-learning algorithms for pigmented skin lesion classification: an open, web-based, international, diagnostic study.
] and most recently, our group demonstrated that the combination of human and artificial intelligence is superior to the individual results of dermatologists and CNNs [
While these developments show great potential to improve the standard of care, 30–50% of all melanomas and more than half of those in young patients evolve from initially benign lesions [
]. Despite its high relevance for patients undergoing skin checks, neither clinicians nor computers are yet able to reliably predict a melanocytic nevus' oncologic transformation. Predicting melanoma in benign nevi would enable earlier detection and enhance prevention. In addition, all currently published algorithms were not trained with prospective data, making it impossible for them to learn image features that predict benign or malignant change. The cause of this lack of diagnostic ability lies in the static nature of lesion presentation in the current standard of care, both for clinicians and algorithms: Image sequences of the same lesion of the same patient are not regularly made at dermatology departments, many of which do not even conduct any photographic documentation of skin lesions (including university hospitals), and those that perform documentation do not take photographs in a sequential fashion or in standardised time frames.
For artificial intelligence algorithms, there are simply not enough labelled prospective training data available. The current literature reveals certain types of melanocytic nevi (i.e. spitzoid/dysplastic nevi) and criteria (i.e. visible vasculature) that, in general, appear to have a higher chance to transform into melanoma, but the agreement between what is considered ‘spitzoid’ and ‘dysplastic’ between clinicians and histopathologists is low, and these criteria do not qualify for a fine-grained assessment. In this connection, two recent studies published in the European Journal of Cancer demonstrate that the precision of pathologists to diagnose melanoma may also be successfully enhanced by artificial intelligence algorithms [
1. Detecting early morphologic footprints of oncogenic mutations
Owing to the cumulative nature of oncogenic mutations in melanocytic nevi, a fine-grained early morphologic footprint should be detectable by an algorithm trained with prospective data: Differentiated melanocytes of the skin are described to have long cell life cycles and are unable to undergo mitosis [
]. Thus, melanocytic nevi may accumulate many mutations during their life cycle which are involved in various signalling cascades regulating proliferation (e.g. BRAF and NRAS), replication (TERT), cell cycle control (CDKN2A), metabolism (PTEN and KIT), apoptosis (TP53), etc. [
], and, conversely, physiologic pro-apoptotic responses to these (immune or intracellular responses) over long periods of time that likely leave morphologic footprints (i.e. small regression zones) which should be detectable in a high-resolution photo of the lesion.
2. Evidence on the importance of E for evolution in the ABCDE rule
] first introduced the ABCD rule for dermoscopic melanoma detection which was based on four criteria which all demonstrated to increase the accuracy of dermoscopy: asymmetry, border, colour and diameter.
Since 1999, this method has been extended by a further criterion by Kittler et al. [
]: E for evolution. The authors incorporated information about morphologic changes of the lesion over time and demonstrated that the frequency of reported changes was significantly higher for melanomas than benign melanocytic nevi. Follow-up investigations confirmed that the clinical evolution is a crucial, if not one of the most important criteria when diagnosing skin lesions [
Impact of dermoscopy and short-term sequential digital dermoscopy imaging for the management of pigmented lesions in primary care: a sequential intervention trial.
Results from an observational trial: digital epiluminescence microscopy follow-up of atypical nevi increases the sensitivity and the chance of success of conventional dermoscopy in detecting melanoma.
]. However, owing to the high amount of patients an individual dermatologist sees, it is unlikely that he or she will remember the exact appearance of the lesion at the preluding skin check. A comparison of images of individual moles over time is already possible in clinical practice with applications offered by different software companies; however, they are not regularly used by clinicians, are not connected with prospective research efforts and come with considerable costs for the patients. In addition, most patients, especially those who are at risk for nevus-associated melanoma (i.e. with multiple melanocytic nevi) [
], will not necessarily remember if a lesion has changed or not. Furthermore, neither clinician nor current algorithms can fully assess whether a change itself was of malignant or benign character.
3. Enhancing the E of the ABCDE rule with prospective data sets
An algorithm trained on large data sets of prospective images of nevi remaining nevi and nevi transforming into melanoma over long periods of time would have the following predictive abilities:
a)
Detect a morphologic footprint which reveals the probability of future oncologic transformation of a nevus.
b)
If two images of the same nevus at two time points are available, assess the change of a nevus itself as rather benign or malignant.
c)
Precisely reveal the morphologic features predicting an oncogenic transformation in an understandable manner by the means of modern methods for interpreting and understanding deep neural networks [
All these aspects go beyond what is currently possible.
4. How to generate prospective data sets?
To yield the proposed functionality of machine learning, sequential images of many individual lesions from individual patients with follow-up images over long time frames (i.e. months to years) are necessary. Mole-monitoring apps which acquire such prospective data by mapping moles over long periods of time via smartphone cameras have been around for years. However, the primary focus of these apps is to help smartphone users to monitor moles, but to date, no app that generates prospective image data specifically for research is available. Thus, the development of such a free research-based mole-monitoring app was initiated at the medical campus in Heidelberg together with colleagues from the Department of Dermatology at the University Hospital of Essen (Fig. 1; mock-ups of the apps prototype).
Fig. 1Features of the plannedmole-monitoring app. Users can generate high-resolution images from specific distances (left), taken in specific time frames (middle), of the same mole(s) with a pin-pointed body location over a long period of time (right). All image sets are made exportable via an admin panel and are annotated by time frame, exact location and user ID. The app informs the patient that it is not a replacement for a visit at the physician but a help to self-monitor skin lesions which always should be shown to a dermatologist in case they change over time.
The app will enable to export prospective user-generated data sets for the training of an algorithm based on informed consent. User-generated sequential image sets will in turn be labelled with the precision by artificial intelligence algorithms trained on biopsy-verified images (and thereby exceeding dermatologists’ precision) to be used for supervised learning of a deep neural network with the previously discussed abilities. An alternative would be to have the user self-label images after a biopsy was performed. Although this option will be offered within the app, it is deemed to be less reliable by the investigators because of a lack of compliance.
Of the vast majority of melanocytic nevi, only very few nevi will develop into melanoma [
], and this in turn is time dependent. Thus, the achievement of a large-enough data set (estimated to be at least 2000 user-generated nevi data sets with a combined follow-up period of approximately 1000 years for first assessments) with current data extraction strategies and based on past experiences of user acquisition by research-based apps launched by the investigators will take roughly three years [
]. Alternative strategies associated with substantially higher costs and efforts, however, would be the change of the current standard of care by enforcing digital follow-up images at standardised time intervals in dermatology departments.
5. The clinical value of melanoma prediction in benign nevi
While the overall prevalence of nevus-associated melanomas was described to be about 33% among all age groups [
]. While malignant growth and metastasis can occur within weeks (especially in young patients), the costs for a comprehensive skin cancer screening in Europe are either not covered (i.e. not covered for young patients until the age of 35 years in Germany) or only covered in intervals of multiple years (i.e. every two years for people older than 35 years in Germany). Thus, predictive diagnostics, especially for young patients with numerous nevi, may reduce the number of melanoma-related deaths, specifically by enabling more focused self-monitoring, early removal of nevi with a high probability of transformation but also by enhancing the precision of current melanoma classifiers already increasingly implemented in routine care.
A convolutional neural network trained with dermoscopic images performed on par with 145 dermatologists in a clinical melanoma image classification task.
Comparison of the accuracy of human readers versus machine-learning algorithms for pigmented skin lesion classification: an open, web-based, international, diagnostic study.
Impact of dermoscopy and short-term sequential digital dermoscopy imaging for the management of pigmented lesions in primary care: a sequential intervention trial.
Results from an observational trial: digital epiluminescence microscopy follow-up of atypical nevi increases the sensitivity and the chance of success of conventional dermoscopy in detecting melanoma.
30409-5&id=gr1.jpg){kind=link}