Deep learning-based, computer-aided classifier developed with a small dataset of clinical images surpasses board-certified dermatologists in skin tumor diagnosis.

https://www.ncbi.nlm.nih.gov/pubmed/29953582

Br J Dermatol. 2018 Jun 28. doi: 10.1111/bjd.16924. [Epub ahead of print]

Deep learning-based, computer-aided classifier developed with a small dataset of clinical images surpasses board-certified dermatologists in skin tumor diagnosis.

Fujisawa Y1, Otomo Y2, Ogata Y3, Nakamura Y1, Fujita R2, Ishitsuka Y1, Watanabe R1, Okiyama N1, Ohara K4, Fujimoto M1.

Author information

1

Dermatology Division, University of Tsukuba.

2

Kyocera Communication System Co., Ltd.

3

KCCS Mobile Engineering Co., Ltd.

4

Dermatology, Akasaka Toranomon Clinic.

Abstract

BACKGROUND:

Application of deep-learning technology to skin cancer classification can potentially improve skin cancer screening sensitivity and specificity, but the number of training images required for such system is thought to be extremely large.

OBJECTIVE:

To determine if deep-learning technology could be used to develop an efficient skin cancer classifying 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 tumors 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 9 dermatology trainees.

RESULTS:

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 < .01), the DCNN achieved greater accuracy, as high as 92.4% ± 2.1% (P < .0001).

CONCLUSIONS:

We developed an efficient skin tumor classifier using a DCNN trained on a relatively small dataset. The DCNN could classify images of skin tumors more accurately than board-certified dermatologists. Collectively, the current system may have capabilities for screening purposes in general medical practice, particularly because it only requires a single clinical image for classification. This article is protected by copyright. All rights reserved.

This article is protected by copyright. All rights reserved.

PMID:

 

29953582

 

DOI:

 

10.1111/bjd.16924

영국 피부과 저널

산학 협동 논문

Deep CNN

민감도 96.3%   특이도  89.5%

피부과 수련의 (74.4%) 피부과 전문의 (85.3%) , DCNN (92.4%) ...

https://onlinelibrary.wiley.com/doi/abs/10.1111/bjd.16924