基于深度学习的口腔鳞癌的诊断与分割方法

DIAGNOSIS AND SEGMENTATION METHODS OF ORAL SQUAMOUS CELL CARCINOMA BASED ON DEEP LEARNING

针对实际应用中要对图片分类并对癌变图进行癌变区域定位的需求,收集天津市口腔医院典型病例,建立口腔细胞病理切片图像数据集,提出基于深度学习的诊断与分割方法。采用以DenseNet为架构的卷积神经网络对图像进行正常与癌变的分类,利用图像分块思想对高分辨率图像分块进行训练,采用迁移学习和数据增强方法减少过拟合问题的发生。分类完成后,使用以DenseNet网络作为编码结构的UNet++分割网络对判断为癌变的图像进行癌变区域定位,采用组合交叉熵方法确定损失函数进行调优。实验表明,该方法能够较好地完成口腔细胞切片图像的分类识别,识别准确率达98.46%,与金标准对比,得到了较理想的分割结果。该方法有助于口腔细胞病理自动诊断系统的开发,可用于口腔鳞癌病理辅助诊断。

In view of the need to classify images and locate the cancerous region of cancerous images in practical applications, the typical cases of Tianjin Stomatology Hospital were collected, the image data set of oral cell pathology slices was established, and the diagnosis and segmentation method based on deep learning was proposed. The convolutional neural network based on DenseNet was used to classify the images into normal and canceration, and the high-resolution images were divided into patches, and the migration learning and data enhancement methods were used to reduce the occurrence of overfitting. After the classification, the UNet++ segmentation network with DenseNet network as the coding structure was used to locate the cancerous area of the image, and the combined cross entropy method was used to determine the loss function for tuning. The experimental results show that this method can complete the classification and recognition of oral cell slice images with a recognition accuracy of 98.46%, compared with the gold standard, a better segmentation result was obtained. This method is helpful for the development of automatic pathological diagnosis system of oral cell pathological and can be used for the auxiliary diagnosis of oral squamous cell carcinoma.