多分支压缩和激励网络的蕈样肉芽肿细胞分类

Classification of Mycosis Fungoides Cells Based on Multi Branch Squeeze and Excitation Network

基于中国医学科学院皮肤病研究所收集的77张早、中期皮肤蕈样肉芽肿全扫描切片,构建了多分支压缩和激励网络模型,实现了皮肤蕈样肉芽肿的淋巴样细胞与上皮细胞的分类。模型分为编码和解码2个阶段,编码阶段对应一个分支;解码阶段有3个分支,对应一个主任务和2个辅助任务。主任务分支输出细胞分类的结果,辅助分支I输出细胞与背景,辅助分支II输出水平垂直边界图谱。在训练阶段,从切片中选取576张图像块,由专业病理科医生进行标记,其中将464张图像块用于训练,将112张图像块用于验证,最后在全扫描切片上进行测试。模型的细胞分割准确率为0.943,F1值为0.728,细胞的分类平均准确率为0.943。实验结果表明,所提出的模型能够实现皮肤蕈样肉芽肿淋巴样细胞和上皮细胞的识别与分类,为皮肤蕈样肉芽肿的计算机辅助诊断奠定了重要基础。

To study the different cell components of mycosis fungoides, a multi branch squeeze and excitation network model is constructed based on 77 whole slide images of early and middle stage mycosis fungoides, and the classification of lymphocytes and epithelial cells of mycosis fungoides is realized. The network is divided into two stages: encoding and decoding. The encoding stage corresponds to one branch, and the decoding stage has three branches, corresponding to one main task and two auxiliary tasks. The main task branch outputs the results of cell classification, the auxiliary branch I outputs the cells and background, and the auxiliary branch II outputs the horizontal and vertical boundary map. In the training stage, 576 image blocks were selected from the slices and marked by professional pathologists, including 464 for training and 112 for verification. Finally, they are tested on the whole slide images. The cell segmentation accuracy and F1 score of the model are 0.943 and 0.728, respectively. The average accuracy of classification is 0.943. The experimental results show that the proposed model can recognize and classify lymphocytes and epithelial cells in mycosis fungoides, which lays an important foundation for computer-aided diagnosis of cutaneous mycosis fungoides.