基于深度级联网络的乳腺淋巴结全景图像的癌转移区域自动识别

淋巴结癌转移区域的自动识别是乳腺癌病理分期的重要前提。但由于全景图像尺寸巨大, 组织形态复杂多样, 在乳腺淋巴结全景图像中自动检测和定位癌转移区域具有很大的难度。设计一种基于深度级联网络的方法, 实现对乳腺淋巴结全景图像癌转移区域的自动定位与识别。采用由粗定位到精定位的两个深度网络模型级联的方式, 首先基于医生标记的癌转移区域, 提取阳性与阴性图像块训练粗定位网络VGG16得到粗定位结果, 然后对比粗定位结果与医生标记提取阳性和假阳性区域的图像块, 再训练精定位的ResNet50网络用于识别阳性和假阳性区域。为了验证所提出深度级联网络的有效性, 选用Camelyon16公开的共400张乳腺淋巴结全景图像数据集用作训练和测试。结果表明, 所提出的VGG16+ResNet50级联网络模型的定位指标FROC得分达到0.891 2, 分别比单个深度网络模型VGG16和ResNet50的FROC得分高0.153 1和0.147 0, 比AlexNet+VGG16级联的网络模型FROC得分高0.028 8, 显示深度级联网络模型对淋巴结癌转移区域可以实现更加精准的识别。

Deep Cascaded Network for Automated Detection of Cancer MetastasisRegion from Whole Slide Image of Breast Lymph Node

Automated recognition of the cancer metastasis region in lymph nodes is an essential prerequisite for the pathological staging of breast cancer. However, due to the massive size of panoramic images and the complexity and diversity of tissue morphology, it is challenging to automatically detect and locate the cancer metastasis areas in panoramic images of the lymph nodes. In this paper, a method based on the deep cascaded network was proposed to realize the automatic localization and recognition of tumor metastasis region in panoramic images of breast lymph nodes. We implemented a coarse-to-fine model cascading method and a coarse positioning network VGG16 was first trained based on positive and negative image blocks extracted from doctor marked region, and then compared with the doctor marked region to extract the image blocks from the positive and false positive areas. The finely positioned ResNet50 network was trained to identify the positive and false-positive regions. The effectiveness of the deep cascaded network was verified with a Camelyon16 dataset, which included a total of 400 whole slide images for training and testing. The FROC value of the positioning index of the VGG16+ResNet50 cascaded network model proposed in this paper reached 0.891 2, which was 0.153 1 and 0.147 0 higher than the single deep network models VGG16 and ResNet50, and only 0.028 8 higher than AlexNet+VGG16 cascaded network model, showing that the deep cascaded network model could achieve more accurate identification of lymph node cancer metastasis regions.