AUTHOR=Yin Wenjun , Huang Jianhua , Chen Jianlin , Ji Yuanfa 

TITLE=A study on skin tumor classification based on dense convolutional networks with fused metadata

JOURNAL=Frontiers in Oncology

VOLUME=Volume 12 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2022.989894

DOI=10.3389/fonc.2022.989894

ISSN=2234-943X

ABSTRACT=Skin cancer is the most common and lethal cancer in human existence, and the diagnostic accuracy of experienced expert physicians is only about 60% according to statistics. In order to improve the accuracy of skin cancer diagnosis, a skin tumor classification model incorporating dense convolutional networks with skin tumor clinical patient metadata is proposed. Firstly, the pre-trained model of DenseNet-169 on ImageNet dataset is trained on skin tumor dataset by tuning the reference and extracting the high level features implied by the images; secondly, the MetaNet module is introduced to control a specific part of each feature channel of the DenseNet-169 network by metadata to obtain the weighted features; meanwhile, the MetaBlock module is introduced to enhance the features extracted from the image by using metadata, i.e., to guide the image to select the most relevant features for output based on the metadata information; finally, the MD-Net（Metadata Networks） module is constructed by fusing the features of the MetaNet and MetaBlock modules through dimensionality reduction and expansion operations, and is input to the classifier to obtain the classification results. The experimental results show that the DenseNet network model fused with this module achieves 81.4% in the balance accuracy index, and the correct diagnosis rate is improved by about 8%~15.6% compared with the accuracy of the existing work, and solves the problem that a few skin tumor categories are not diagnosed with high accuracy.