Dermatologist-level classification of skin cancer with deep neural networks

We used the lesional steps in tumor progression and multivariable logistic regression to develop a prognostic model for primary, clinical stage I cutaneous melanoma. This model is 89% accurate in predicting survival. Using histologic criteria, we assigned melanomas to tumor progression steps by ascertaining their particular growth phase. These phases were the in situ and invasive radial growth phase and the vertical growth phase (the focal formation of a dermal tumor nodule or dermal tumor plaque within the radial growth phase or such dermal growth without an evident radial growth phase). After a minimum follow-up of 100.6 months and a median follow-up of 150.2 months, 122 invasive radial-growth-phase tumors were found to be without metastases. Eight-year survival among the 264 patients whose tumors had entered the vertical growth phase was 71.2%. Survival prediction in these patients was enhanced by the use of a multivariable logistic regression model. Twenty-three attributes were tested for entry into this model. Six had independently predictive prognostic information: (a) mitotic rate per square millimeter, (b) tumor-infiltrating lymphocytes, (c) tumor thickness, (d) anatomic site of primary melanoma, (e) sex of the patient, and (f) histologic regression. When mitotic rate per square millimeter, tumor-infiltrating lymphocytes, primary site, sex, and histologic regression are added to a logistic regression model containing tumor thickness alone, they are independent predictors of 8-year survival (P less than .0005).

Image-based computer aided diagnosis systems have significant potential for screening and early detection of malignant melanoma. We review the state of the art in these systems and examine current practices, problems, and prospects of image acquisition, pre-processing, segmentation, feature extraction and selection, and classification of dermoscopic images. This paper reports statistics and results from the most important implementations reported to date. We compared the performance of several classifiers specifically developed for skin lesion diagnosis and discussed the corresponding findings. Whenever available, indication of various conditions that affect the technique's performance is reported. We suggest a framework for comparative assessment of skin cancer diagnostic models and review the results based on these models. The deficiencies in some of the existing studies are highlighted and suggestions for future research are provided.