Abstract:

Skin cancer is one of the most common cancers in humans.However, its diagnosis is highly influenced by individual physicians' subjectivity, and most existing studies on skin cancer diagnosis by neural networks stay at the image level without considering the clinical data of patients.Thus, its diagnostic accuracy must be improved.An MetaData Layer(MD-Layer) module that fuses the metadata of clinical patients with skin tumors is proposed and embedded into a dense convolutional classification network model DenseNet-169.The DenseNet-169 network is pre-trained on the ImageNet data set, and the obtained model is trained on the skin cancer data set to extract high-level features hidden in the image. The MD-Layer module is constructed by fusing the features extracted by the MetaNet and MetaBlock modules. The MetaNet module controls specific parts of each feature channel in the DenseNet-169 network through metadata to obtain weighted features.The MetaBlock module uses metadata to enhance the features extracted from the image, that is, guides the image to select the most relevant feature output according to the metadata information.Finally, the fused results are input to the classifier to realize the classification of skin tumors.The experimental results show that the DenseNet-169 network model fused with the MD-Layer module achieves an 0.814 Balance Accuracy(BACC) index, and the BACC is improved by approximately 0.080-0.156 compared with existing approaches while solving the problem of low diagnostic accuracy for a few skin tumor categories.

Key words: skin tumor classification, DenseNet-169 model, metadata, feature fusion, MD-Layer module

摘要：

皮肤癌的诊断受医生个人主观影响较大，现有的神经网络皮肤癌诊断研究大多停留在图像层面，没有考虑患者的临床数据，诊断准确率有待提高。提出一种融合皮肤肿瘤临床患者元数据的MD-Layer模块，并嵌入到密集卷积分类网络模型DenseNet-169中。在ImageNet数据集上对DenseNet-169网络进行预训练，使得到的模型在皮肤癌数据集上进行调参训练，并提取图像隐含的高层次特征。MD-Layer模块由MetaNet模块和MetaBlock模块提取到的特征融合构建而来。MetaNet模块通过元数据控制DenseNet-169网络中每个特征通道的特定部分，从而获得加权特征。MetaBlock模块利用元数据增强从图像中提取的特征，即根据元数据信息引导图像选择最相关的特征输出。最后，将融合后的结果输入到分类器，实现皮肤肿瘤的分类。实验结果表明，融合MD-Layer模块的DenseNet-169网络模型的平衡准确率为0.814，相较于已有工作提升0.080~0.156，解决了少数皮肤肿瘤类别诊断准确率不高的问题。

关键词: 皮肤肿瘤分类, DenseNet-169模型, 元数据, 特征融合, MD-Layer模块