Premium
Deep‐learning‐based, computer‐aided classifier developed with a small dataset of clinical images surpasses board‐certified dermatologists in skin tumour diagnosis
Author(s) -
Fujisawa Y.,
Otomo Y.,
Ogata Y.,
Nakamura Y.,
Fujita R.,
Ishitsuka Y.,
Watanabe R.,
Okiyama N.,
Ohara K.,
Fujimoto M.
Publication year - 2019
Publication title -
british journal of dermatology
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 2.304
H-Index - 179
eISSN - 1365-2133
pISSN - 0007-0963
DOI - 10.1111/bjd.16924
Subject(s) - medicine , medical diagnosis , convolutional neural network , artificial intelligence , classifier (uml) , deep learning , skin cancer , clinical practice , diagnostic accuracy , certification , medical physics , radiology , cancer , computer science , family medicine , political science , law
Summary Background Application of deep‐learning technology to skin cancer classification can potentially improve the sensitivity and specificity of skin cancer screening, but the number of training images required for such a system is thought to be extremely large. Objectives To determine whether deep‐learning technology could be used to develop an efficient skin cancer classification system with a relatively small dataset of clinical images. Methods A deep convolutional neural network ( DCNN ) was trained using a dataset of 4867 clinical images obtained from 1842 patients diagnosed with skin tumours at the University of Tsukuba Hospital from 2003 to 2016. The images consisted of 14 diagnoses, including both malignant and benign conditions. Its performance was tested against 13 board‐certified dermatologists and nine dermatology trainees. Results The overall classification accuracy of the trained DCNN was 76·5%. The DCNN achieved 96·3% sensitivity (correctly classified malignant as malignant) and 89·5% specificity (correctly classified benign as benign). Although the accuracy of malignant or benign classification by the board‐certified dermatologists was statistically higher than that of the dermatology trainees (85·3% ± 3·7% and 74·4% ± 6·8%, P < 0·01), the DCNN achieved even greater accuracy, as high as 92·4% ± 2·1% ( P < 0·001). Conclusions We have developed an efficient skin tumour classifier using a DCNN trained on a relatively small dataset. The DCNN classified images of skin tumours more accurately than board‐certified dermatologists. Collectively, the current system may have capabilities for screening purposes in general medical practice, particularly because it requires only a single clinical image for classification.