摘要
目的基于DeepLab V3+网络框架,构建深度学习计算机视觉模型,实现对上消化道溃疡内镜图片较精确的语义分割。方法DeepLab V3+网络框架的编码器首先通过多个并行的、具有不同采样率的空洞卷积层,之后通过一个全局平均池化层对图像进行特征提取,实现多尺度提取特征;而解码器部分对深度特征层进行4次上采样,与浅层特征层进行堆叠并调整堆叠层大小,使其和输入图像大小一致,得到模型的预测结果。结果在内部验证集中,该模型的准确度(Accuracy,ACC)为0.963,平均交并比(Mean Intersection Over Union,mIoU)为0.927;外部测试集中,该模型的ACC为0.958,mIoU为0.915;均优于传统算法U-Net(内部验证集ACC为0.810,mIoU为0.785;外部测试集ACC为0.779,mIoU为0.732)。结论DeepLab V3+网络框架在识别病灶方面准确度高,具有较好的临床实践性。
Objective Based on DeepLab V3+network framework,to construct a deep learning computer vision model to achieve more accurate semantic segmentation of upper gastrointestinal ulcer endoscopic images.Methods The encoder of the DeepLab V3+network framework initially extracted features from images using multiple parallel dilated convolutional layers with varying sampling rates,followed by global average pooling to achieve multi-scale feature extraction.In the decoder section,the deep feature layer was up-sampled four times while the shallow feature layer was stacked and adjusted in size to match that of the input image before generating predictions.The prediction results of the model were obtained.Results In the internal validation set,the accuracy(ACC)of the model was 0.963,and the mean intersection over union(mIoU)was 0.927.In the external test set,the ACC and mIoU were 0.958 and 0.915.All of them were better than the traditional algorithm U-Net(internal validation set ACC was 0.810,mIoU was 0.785;the external test set ACC was 0.779 and mIoU was 0.732).Conclusion The DeepLab V3+network framework has high accuracy in identifying lesions,and has good clinical practice.
作者
薛雨涵
周亦佳
何宇
林嘉希
朱锦舟
刘晓琳
王玉
许春芳
殷民月
XUE Yuhan;ZHOU Yijia;HE Yu;LIN Jiaxi;ZHU Jinzhou;LIU Xiaolin;WANG Yu;XU Chunfang;YIN Minyue(Department of Gastroenterology,The First Affiliated Hospital of Soochow University,Suzhou Jiangsu 215006,China;Suzhou Medical College of Soochow University,Suzhou Jiangsu 215123,China;Suzhou Clinical Center of Digestive Diseases,Suzhou Jiangsu 215006,China;Department of General Surgery,Jintan Hospital Affiliated to Jiangsu University,Changzhou Jiangsu 213200,China)
出处
《中国医疗设备》
2023年第11期22-26,共5页
China Medical Devices
基金
国家自然科学基金(81900508,82000540)
江苏省自然科学基金(BK20190172)
苏州市科技计划(SKY2021038)
苏州市科教兴卫项目(KJXW2019001)
苏州大学医学部学生课外科研项目(2021YXBKWKY050)。
