Abstract:

Time-consuming artificial diagnosis of pathological images will cause visual fatigue of doctors, while both misdiagnosis and missed diagnosis are easy to occur. In response to the above phenomena, a method combining multidimensional features of convolutional neural network is proposed to quickly and accurately identify lesion in pathological images. ROI extraction and image cutting is used to obtain small-scale block data. The method of stain correction is used to solve the problems of uneven staining and weak contrast in block data. A deep learning model is built, using several depthwise separable convolution to extract features of different dimensions, adding residual connection to avoid gradient disappearance, combining the feature information of different dimensions to improve feature utilization. The experimental results show that stain correction can improve prediction accuracy and the above model has the characteristics of few parameters and strong robustness. At the same time, the accuracy of lesion recognition in pathological images can reach a high level, while both false positive rate and false negative rate are low, so it will have a broad application prospect in the future.

Key words: multidimensional, deep learning, convolutional neural network, depthwise separable, stain correction, pathological image, lesion

摘要：

长时间的病理图像人工诊断会使医生产生视觉疲劳，误诊和漏诊情况容易发生。针对以上现象，提出一种结合卷积神经网络中多维度特征的方法，快速准确识别出病理图像中的病灶区域。使用感兴趣区提取及图像裁剪获得小尺寸图块数据；使用染色校正的方法以解决图块染色不均，对比度弱等问题；搭建深度学习模型，使用多组深度可分离卷积提取不同尺度的特征，加入残差连接以避免梯度消失，联合不同维度的特征信息以提高特征利用率。实验结果表明，染色校正能够提高预测准确率，上述模型具有参数少、鲁棒性强的特点，最终对病理图像病灶的识别均能达到较高的准确率，假阳性及假阴性均较低，未来将具有广泛的应用前景。

关键词: 多维度, 深度学习, 卷积神经网络, 深度可分离, 染色校正, 病理图像, 病灶