融合迁移学习和数据增强的SC-Net模型在皮肤癌识别中的应用

为了解决皮肤癌诊断模型中性能无法满足临床应用要求，对于少数类别诊断精度不高的问题，提出一种基于迁移学习和数据增强的皮肤癌诊断模型SC-Net（skin cancer-net）。首先，引入ECA注意力模块，把DenseNet-201在ImageNet数据集上的预训练模型在皮肤癌数据集上进行微调训练并提取图像隐含高层次特征；然后融合一般性统计特征，并且通过SMOTE过采样技术以增强少数类别数据；最后，将数据输入XGBoost模型进行训练，最终得到SC-Net分类模型。实验结果表明，SC-Net模型在准确率、灵敏度、特异度三个指标上达到99.25%、99.25%和99.88%，诊断准确率相对于已有文献精度提升约0.6%~18.7%，并且对于皮肤纤维瘤、光化性角化病等少数类别具备更强的分类能力。

Application of SC-Net model integrated with transfer learning and data augmentation in skin cancer classification

The performance of the current skin cancer diagnostic models cannot meet the requirements of clinical applications, and the diagnostic accuracy is not high for a few categories. To slove this problem, this paper proposed a SC-Net model based on transfer learing and data augmentation. Firstly, it used the ECA attention module to fine tune the pre-training model of DenseNet-201 on the skin cancer dataset, and extracted the implicit high-level features of the images. Then, it joined the general statistical features, and used SMOTE oversampling technology to balance a few categories of data. Finally, it putt the data into XGBoost model for training to obtain the final SC-Net classification model. The experimental results show that the accuracy, sensitivity and specificity of SC-Net model reach 99.25%, 99.25% and 99.88%, which is about 0.6%~18.7% higher than the existing models. The proposed model has stronger classification ability for a few categories such as Dermato fibroma and Actinic keratoses and intraepithelial carcinoma.