Automated image classification by pattern recognition is a branch of machine learning (ML) which offers the dermatologist a useful tool for assessment in the diagnosis of skin cancer.1 Deep convolutional neural networks (DCNN) have dramatically improved accuracy in feature learning and object classification2 and have been successfully used in the classification of dermoscopic images of skin lesions.3 However, the selection of images may include certain special features which prevent its universal use at the present time. In this study we assessed some exclusion criteria in the selection of skin cancer images (with an emphasis on melanoma) for ML analysis, according to recent works in this field.1,4,5

This study was conducted in a tertiary academic skin cancer center in Barcelona, Spain. A retrospective cohort study was designed including 2,849 consecutive high-quality dermoscopy images of skin tumors from the Melanoma Unit database from 2010 to 2014. The DermLite® photo digital epiluminescence microscopy system 3Gen with 37mm thread size and a Canon camera, model G16 were used. Pathological diagnosis was available for 2,429 images. Finally, the images were assorted according to their theoretical eligibility for ML analysis, pursuant to some potential exclusion criteria1,4,5: difficulty in lesion border detection (absence of pigmentation, absence of normal surrounding skin, presence of hair, location on volar skin), metastasis or an ulcerated lesion.

This study has been approved by the institutional review board. All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional research committee and with the 1964 Helsinki declaration and its later amendments or comparable ethical standards.

Out of the 2,849 images from our database, 968 (34%) were selectable as they did not have any potential exclusion criteria for analysis by a ML system. Nevi, melanoma and basal cell carcinoma were the most frequent lesions in our database. Only 64.7% of nevi and 36.6% of melanoma did not have any potential exclusion criteria (Table 1).

Of 528 melanomas, 335 (63.4%) could potentially be excluded. An absence of normal surrounding skin (40.5% of all melanomas) and absence of pigmentation (14.2%) were the most common reasons for exclusion from ML analysis. Other reasons for exclusion are shown in Table 1.

Melanoma accounts for the majority of skin cancer deaths. Early diagnosis and treatment significantly improves its prognosis. The development of an effective screening method is needed and automated image classification by pattern recognition may achieve diagnostic accuracy similar to expert dermatologist.6 However, some limitations have to be overcome. One of these is the exclusion criteria in the selection of skin cancer images. While solely high-quality dermoscopy images were selected from our database, only 34% did not have any potential exclusion criteria for classification by most state-of-the-art ML algorithms. Moreover, 63.4% of our melanomas had at least one of the potential exclusion criteria mentioned above. This considerably decreases diagnostic accuracy and utility of some ML systems. Large lesions are a serious problem for ML algorithms, as they do not fit within the diameter of the majority of dermoscopy lenses, and this renders all the state-of-the-art systems which need to pre-compute lesion segmentation.1 Even if some works have proposed hair detection/removal methods,5 most ML systems’ performance is deteriorated by its presence. Since most dermoscopy datasets for algorithm training don’t include volar skin lesions, the systems trained on these won’t be able to correctly classify acral lesions. Nevertheless, the artificial intelligence community is rapidly moving to overcome these nuances. Yu et al.7 published recently a work where DCNN was used for acral melanoma and nevus classification. In this work we consider the limitations of most but not all ML systems.

Our study shows that the main potential exclusion criteria were the absence of normal surrounding skin and the absence of pigmentation. Many melanomas developed in sun-damaged skin with abnormal surrounding skin, which makes them unsuitable for analysis by most of the current ML systems due to difficulties in lesion border detection.5 Moreover, amelanotic melanoma which accounts for 2%–8% of all melanomas8 cannot yet be diagnosed by most current ML systems. This could be addressed by designing ML systems which are able to work with images which do not contain the entire lesion and increasing the dataset size, selecting a higher number of representative dermoscopy images.

In conclusion, we consider that ML systems, especially those based in the new developments in the deep learning field will not only convert ML into a valuable tool for the dermatologist but also for the general population. However, these systems are able to overcome some limitations to enlarge spectrum of measurable images. It is clear though that researchers are moving forward towards this direction, since some of the exclusion criteria mentioned in this work have already been overcome by recent algorithms included in the ISIC International Symposium.3

The study in the Melanoma Unit, Hospital Clínic, Barcelona was supported in part by grants from Fondo de Investigaciones Sanitarias P.I. 12/00840, PI15/00956 and PI15/00716 Spain; by the CIBER de Enfermedades Raras of the Instituto de Salud Carlos III, Spain, co-funded by “Fondo Europeo de Desarrollo Regional (FEDER). Unión Europea. Una manera de hacer Europa”; by the AGAUR 2014_SGR_603 and 2017_SGR_1134 of the Catalan Government, Spain; by a grant from “Fundació La Marató de TV3, 201331-30”, Catalonia, Spain; by the European Commission under the 6th Framework Programme, Contract n°: LSHC-CT-2006-018702 (GenoMEL); by CERCA Programme/Generalitat de Catalunya and by a Research Grant from “Fundación Científica de la Asociación Española Contra el Cáncer” GCB15152978SOEN, Spain. Part of the work was developed at the building Centro Esther Koplowitz, Barcelona.

The authors declare that they have no conflicts of interest