Clinical image identification of basal cell carcinoma and pigmented nevi based on convolutional neural network.
10.11817/j.issn.1672-7347.2019.190205
- Author:
Bin XIE
1
,
2
;
Xiaoyu HE
3
;
Weihong HUANG
4
;
Minxue SHEN
5
,
6
,
7
;
Fangfang LI
5
,
6
,
7
;
Shuang ZHAO
5
,
6
,
7
Author Information
1. School of Automation, Central South University, Changsha 410083
2. Mobile Health Ministry of Education-China Mobile Joint Laboratory, Xiangya Hospital, Central South University, Changsha 410008, China.
3. School of Automation, Central South University, Changsha 410083, China.
4. Mobile Health Ministry of Education-China Mobile Joint Laboratory, Xiangya Hospital, Central South University, Changsha 410008, China.
5. Department of Dermatology, Xiangya Hospital, Central South University
6. Hunan Key Laboratory of Skin Cancer and Psoriasis
7. Hunan Engineering Research Center of Skin Health and Disease, Changsha 410008, China.
- Publication Type:Journal Article
- MeSH:
Area Under Curve;
Carcinoma, Basal Cell;
diagnostic imaging;
Humans;
Neural Networks, Computer;
Nevus, Pigmented;
diagnostic imaging;
Skin Neoplasms;
diagnostic imaging
- From:
Journal of Central South University(Medical Sciences)
2019;44(9):1063-1070
- CountryChina
- Language:Chinese
-
Abstract:
To construct an intelligent assistant diagnosis model based on the clinical images of basal cell carcinoma (BCC) and pigmented nevi in Chinese by using the advanced convolutional neural network (CNN).
Methods: Based on the Xiangya Medical Big Data Platform, we constructed a large-scale clinical image dataset of skin diseases according to Chinese ethnicity and the Xiangya Skin Disease Dataset. We evaluated the performance of 5 mainstream CNN models (ResNet50, InceptionV3, InceptionResNetV2, DenseNet121, and Xception) on a subset of BCC and pigmented nevi of this dataset. We also analyzed the basis of the diagnosis results in the form of heatmaps. We compared the optimal CNN classification model with 30 professional dermatologists.
Results: The Xiangya Skin Disease Dataset contains 150 223 clinical images with lesion annotations, covering 543 skin diseases, and each image in the dataset contains support for pathological gold standards and the patient's overall medical history. On the test set of 349 BCC and 497 pigmented nevi, the optimal CNN model was Xception, and its classification accuracy can reach 93.5%, of which the area under curve (AUC) values were 0.974 and 0.969, respectively. The results of the heatmap showed that the CNN model can indeed learn the characteristics associated with disease identification. The ability of the Xception model to identify clinical images of BCC and Nevi was basically comparable to that of professional dermatologists.
Conclusion: This study is the first assistant diagnosis study for skin tumor based on Chinese ethnic clinical dataset. It proves that CNN model has the ability to distinguish between Chinese ethnicity's BCC and Nevi, and lays a solid foundation for the following application of artificial intelligence in the diagnosis and treatment for skin tumors.
