BACKGROUND:Smartphone applications ('apps') that use artificial intelligence (AI) for skin cancer diagnosis may help with early detection. However, prospective studies assessing real-world performance are scarce. OBJECTIVES:To evaluate in an independent prospective study the diagnostic accuracy of a widely used skin cancer detection app and the effect of photographic conditions in a user-representative cohort of patients consulting for lesions of concern. METHODS:Between 1 February 2021 and 30 June 2023, participants presenting at an early-access consultation with a lesion of concern were consecutively enrolled. Following dermatological assessment, each lesion was photographed using the app. The app's convolutional neural network (CNN) generates a binary risk output and advice. A teledermatology review was performed in a subset of cases. The diagnostic accuracy of the CNN combined with teledermatology review was compared with final clinical or histopathological diagnosis. Performance was tested under different photographic conditions (angle, lighting, user) and with different smartphone models. The trial was registered with ClinicalTrials.gov (NCT05246163). RESULTS:A total of 1458 participants with 1904 lesions of concern were included. Of these, 185 (9.7%) were skin cancers, with 32 melanomas. Image capture was unsuccessful in 16.6% of lesions (n = 317/1904), despite optimal conditions. For successfully captured lesions, the CNN achieved a sensitivity of 82.5% and specificity of 76.8% for skin cancer detection. In-app teledermatology review was available for 65.7% (n = 1042/1587) of images. Combined CNN and teledermatology review resulted in increased specificity (86.8%; P < 0.001), with a sensitivity of 75.3%. Overall sensitivity in melanocytic lesions was lower than in nonmelanocytic lesions (71.0% vs. 76.3%; P < 0.001), whereas specificity was higher (90.0% vs. 84.9%; P < 0.001). A substudy demonstrated low image-capture success in user hands (28.9%) and variation in diagnostic performance between smartphone models. CONCLUSIONS:This independent, prospective study evaluated diagnostic performance of a widely used skin cancer detection app in a user-representative cohort. The findings highlight the importance of independent clinical validation of AI-based healthcare tools in real-world settings.

Despite skin toxicity during anti-EGFR therapy occurs frequently and significantly impacts patients’ quality of life, it is often not recognized nor optimally managed. This may lead to unnecessary dose reductions or treatment discontinuations of the cancer treatment. Access to rapid dermatology advice may be limited in hospitals. In this multicenter pilot study, we conclude that teledermatology may add value for both patients and oncologists in the management of treatment-induced skin toxicity.

In dermatology, deep learning may be applied for skin lesion classification. However, for a given input image, a neural network only outputs a label, obtained using the class probabilities, which do not model uncertainty. Our group developed a novel method to quantify uncertainty in stochastic neural networks. In this study, we aimed to train such network for skin lesion classification and evaluate its diagnostic performance and uncertainty, and compare the results to the assessments by a group of dermatologists. By passing duplicates of an image through such a stochastic neural network, we obtained distributions per class, rather than a single probability value. We interpreted the overlap between these distributions as the output uncertainty, where a high overlap indicated a high uncertainty, and vice versa. We had 29 dermatologists diagnose a series of skin lesions and rate their confidence. We compared these results to those of the network. The network achieved a sensitivity and specificity of 50% and 88%, comparable to the average dermatologist (respectively 68% and 73%). Higher confidence/less uncertainty was associated with better diagnostic performance both in the neural network and in dermatologists. We found no correlation between the uncertainty of the neural network and the confidence of dermatologists (R = -0.06, p = 0.77). Dermatologists should not blindly trust the output of a neural network, especially when its uncertainty is high. The addition of an uncertainty score may stimulate the human-computer interaction.

Blue-white veil (BWV) is an important feature in melanoma detection where it is present as a focal, ill-defined area. In benign lesions BWV typically has a more uniform distribution across the lesion. In this work we analyzed spectral signatures, obtained using multispectral dermoscopy, of pathology proven atypical nevi (n = 30) and melanoma (n = 119) lesions using a retrospectively collected dataset. Using AI techniques, we were able to automatically detect the BWV dermoscopic feature. When only considering darker regions in the lesions, we found a specificity in discriminating melanoma from atypical nevi based on the presence of BWV of 100% at a sensitivity of 34%. When additionally including paler regions, the sensitivity increases to 64% at a slightly reduced specificity of 95%. The main advantage of this automatic BWV detection is that it only relies on the observed spectral signature of the corresponding region of the lesion, without requiring spatial information or relative color compared with the surrounding lesion area. Furthermore, the used technique allows to exactly highlight where the BWV was detected, allowing expert review. Future work will include visual analysis by expert dermatologists of the detected BWV areas, especially since literature shows a BWV prevalence of 18.7% in nevi (and 34.9% in melanoma). The higher specificity of the automatic BWV detector may suggest that the presented technique allows to discriminate BWV correlated with benign nevi from BWV correlated with melanoma. Finally, a prospective study should be performed to prove the benefit of the presented automatic BWV detection in clinical practice.

To the Editor: In computer vision, deep learning has achieved excellent performance in image classification. Since then, it has been applied in various medical fields, including dermoscopy. For example, a 2017 study by Esteva et al1Esteva A. Kuprel B. Novoa R.A. et al.Dermatologist-level classification of skin cancer with deep neural networks.Nature. 2017; 542: 115-118https://doi.org/10.1038/nature21056Crossref PubMed Scopus (54) Google Scholar showed that a neural network trained to classify cutaneous tumoral lesions achieved similar performance to trained dermatologists. However, a neural network only outputs a single value per class, the class probability. This value does not reflect the uncertainty and is generally overconfident for one class.2Gal Y, Ghahramani Z. Dropout as a Bayesian approximation: representing model uncertainty in deep learning. In: 33rd International Conference on Machine Learning, ICML 2016. Vol 3. PMLR; 2016:1651-1660.Google Scholar This means that a neural network could output a high class probability while still being uncertain. Unable to reflect uncertainty, an overconfident neural network could instigate a false sense of security. To overcome this, we developed a novel metric for quantifying output uncertainty in stochastic neural networks. These networks no longer output a single probability but a complete distribution per class by analyzing copies of the same images (Fig 1). High overlap between class distributions indicates high uncertainty, while a low overlap indicates low uncertainty.3Van Molle P. Verbelen T. Vankeirsbilck B. et al.Leveraging the Bhattacharyya coefficient for uncertainty quantification in deep neural networks.Neural Comput Appl. 2021; 33: 10259-10275https://doi.org/10.1007/s00521-021-05789-yCrossref Scopus (10) Google Scholar We trained a stochastic neural network (Supplementary material, available via Mendeley at https://data.mendeley.com/datasets/ky4kgp4243/1) using the HAM10000 data set of common pigmented skin lesions4Tschandl P. Rosendahl C. Kittler H. Data descriptor: The HAM10000 dataset, a large collection of multi-source dermatoscopic images of common pigmented skin lesions.Sci Data. 2018; 5https://doi.org/10.1038/sdata.2018.161Crossref PubMed Scopus (949) Google Scholar and evaluated performance using a hold-out test set. The network labeled each lesion and generated a corresponding uncertainty score, achieving a classification accuracy of 80%. We validated our uncertainty metric by rejecting lesions with an uncertainty score above a given threshold. When decreasing the threshold (0.8, 0.6, 0.4, 0.2, and 0.0), the classification accuracy for the remaining lesions was, respectively, 84%, 87%, 90%, 92%, and 96%, demonstrating that a decreasing uncertainty yields higher accuracy. The sensitivity, specificity, and receiver operating characteristics (ROC) area under the curve (AUC) for the management decisions of the network were respectively 63%, 94%, and 0.92. Next, we constructed a subset of 30 lesions, selecting a mix of lesions where the network either produced the correct or wrong label, including samples with high and low uncertainty for both. On this subset, the network scored a classification accuracy of 46%. The sensitivity, specificity, and ROC-AUC curve for the management decisions were, respectively, 50%, 88%, and 0.69. Through an online survey, professional dermatologists scored the same subset. In addition, we prompted their confidence level. In total, this analysis included results from 22 participants. The dermatologists achieved a mean (min-max) classification accuracy of 55% (40%-66%). The mean (min-max) sensitivity, specificity and ROC-AUC for the management decisions were, respectively, 68% (25%-96%), 73% (50%-89%), and 0.70 (0.57-0.85). Individual dermatologists scored similar to the neural network, with the average dermatologist scoring slightly better (Fig 2). When comparing the uncertainty scores of the neural network to the confidence levels of the dermatologists, we found no clear relationship (P > .05). This indicates that both are unsure about different (aspects of a) lesion(s), raising the opportunity to complement each other in diagnosis. In future work, we plan to further focus on the impact of adding uncertainty to neural network-based ranking on dermatologists' diagnosis and management in a real-world setting. None disclosed.

Introduction: While the results of AI algorithms on benchmark challenges seem promising, caution is needed. A lesion inspected in clinical practice is not constrained to the number of lesion types for which the AI was trained. Even when a lesion fits within one of these more prevalent types, its appearance can be atypical, eg, due to different skin type, overlapping scar, etc. Automatically detecting such cases is an important step towards common adaptation of AI. In this experiment we investigate if the features learned by the AI model can help in detecting unsupported lesions.

Background and objectives Dermoscopy has proven its value in the diagnosis of skin cancer and, therefore, is well established in daily dermatology practice. Up until now, analogue white light dermoscopy is the standard. Multispectral dermoscopy is based on illumination of the skin with narrowband light sources with different wavelengths. Each of these wavelengths is differently absorbed by skin chromophores, such as pigment or (de)oxygenated blood. Multispectral dermoscopy could be a way to enhance the visualization of vasculature and pigment. We illustrate possible additional information by such “skin parameter maps” in some cases of basal cell carcinoma and Bowen's disease. Methods Using a new digital multispectral dermatoscope, skin images at multiple wavelengths are collected from different types of skin lesions. These particular images together with the knowledge on skin absorption properties, result in so called “skin parameter maps”. Results A “pigment contrast map,” which shows the relative concentration of primarily pigment, and a “blood contrast map” which shows the relative concentration of primarily blood were created. Especially, the latter is of importance in diagnosing keratinocyte skin cancer hence vascular structures are a characteristic feature, as further illustrated in the study. Conclusions Skin parameter maps based on multispectral images can give better insight in the inner structures of lesions, especially in lesions with characteristic blood vessels such as Bowen's disease and basal cell carcinoma. Skin parameter maps can be used complementary to regular dermoscopy and could potentially facilitate diagnosing skin lesions.

Multiple reports confirm elevated circulating IL-17 levels and increased numbers of Th17 lymphocytes in patients with non-segmental vitiligo. Additionally, melanocyte damaging characteristics have been ascribed to IL-17. A single-arm pilot study using secukinumab in active non-segmental vitiligo was conducted. The large majority of patients developed additional skin depigmentations limiting further enrollment. Overall, laboratory analysis revealed no change in secreted chemokines or Th subsets. Th17 lymphocytes correlated with Th2, Th9, and Th22 cells while an inverse link with Th1 cells and serum sCD25 levels was observed. In contrast, Th17.1 cells correlated positively with Th1 lymphocytes. Confirmatory results were found in an independent group of patients with vitiligo showing a significant increase in Th17.1 and Th1 lymphocytes in progressive vitiligo patients compared to healthy controls, which was not found for Th17 cells. These results do not support a direct pathogenic role of IL-17 or Th17 cells in vitiligo. Nonetheless, a delicate Th17/Th17.1/Th1 balance seems evident which changes markedly according to disease activity. This may offer new treatment options by interfering with cytokines that drive differentiation of Th17 cells toward Th1.

New promising treatments have been developed for psoriasis that target different parts of the interleukin (IL)-23/IL-17 pathway. This approach is believed to be more disease specific, and sparing the T helper 1 pathway might prevent serious long-term adverse events. Moreover, superior Psoriasis Area and Severity Index improvements are observed, which has redefined treatment goals in psoriasis. The new molecules can be divided into different categories, according to the target: blocking agents can target the upstream cytokine IL-23 or the downstream IL-17. In the latter, a variety of targets exist, such as the ligands IL-17A and IL-17F, or a combination thereof, or a subunit of the receptor, IL-17RA. Each target seems to have its own set of advantages and pitfalls, which will impact the treatment decision in clinical practice. In this review, we summarize the current knowledge on the different inhibitors of the IL-23/IL-17 pathway. Furthermore, we briefly discuss the role of IL-17 in other diseases and comorbidities. Finally, we discuss how comprehensive knowledge is needed for the prescribing physician in order to make the most appropriate therapeutic choice for each individual patient.

In computer vision, deep learning has reached state✁of✁the✁art performance in many classification tasks. When it comes to skin lesion classification, deep neural networks are able to match dermatologist performance in a single image classification setting. A shortcoming, however, is that these networks only output the class likelihood, which is no substitute for model (un)certainty. To overcome this, we approximate statistical inference using a technique called “Monte Carlo dropout”, which we extend to obtain a metric for output confidence. We apply this to skin lesion classification.