针对现有皮肤镜图像分割技术分割精度不高的问题,提出了一种基于多尺度信息提取和特征融合的U型网络(multi-scale information extraction and feature fusion U-shaped network,MF-UNet)模型。在U-Net的基础上,在卷积层后加入批归一化...针对现有皮肤镜图像分割技术分割精度不高的问题,提出了一种基于多尺度信息提取和特征融合的U型网络(multi-scale information extraction and feature fusion U-shaped network,MF-UNet)模型。在U-Net的基础上,在卷积层后加入批归一化层,将原本的跳跃连接部分替换为4级特征融合模块,充分利用语义信息和位置信息,在特征提取端末尾加入多尺度空洞卷积模块和多尺度池化模块,增大感受野,利用双路拼接上采样模块进行上采样,减少图像恢复过程中的信息损失。实验表明,相较于U-Net模型,MF-UNet在平均交并比(mean intersection over union,MIoU)上提升了14.32%,在戴斯相似系数(Dice similarity coefficient,DSC)上提升了13.18%,取得了较好的结果。该研究为计算机技术辅助医生进行皮肤病诊断提供了借鉴。展开更多
AIM To evaluate the usefulness of frozen section diagnosis(FSD) of bile duct margins during surgery for extrahepatic cholangiocarcinoma(CCA). METHODS We retrospectively analyzed 74 consecutive patients who underwent s...AIM To evaluate the usefulness of frozen section diagnosis(FSD) of bile duct margins during surgery for extrahepatic cholangiocarcinoma(CCA). METHODS We retrospectively analyzed 74 consecutive patients who underwent surgery for extrahepatic CCA from 2012 to 2017, during which FSD of bile duct margins was performed. They consisted of 40 distant and 34 perihilar CCAs(45 and 55 bile duct margins, respectively). The diagnosis was classified into three categories: negative, borderline(biliary intraepithelial neoplasia-1 and 2, and indefinite for neoplasia), or positive. FSD in the epithelial layer, subepithelial layer, and total layer was compared with corresponding permanent section diagnosis(PSD) postoperatively.Then, association between FSD and local recurrence was analyzed with special reference to borderline.RESULTS Analysis of 100 duct margins revealed that concordance rate between FSD and PSD was 68.0% in the total layer, 69.0% in the epithelial layer, and 98.0% in the subepithelial layer. The extent of remaining biliary epithelium was comparable between FSD and PSD, and more than half of the margins lost > 50% of the entire epithelium, suggesting low quality of the samples. In FSD, the rate of negative margins decreased and that of borderline and positive margins increased according to the extent of the remaining epithelium. Diagnostic discordance between FSD and PSD was observed in 31 epithelial layers and two subepithelial layers. Alteration from borderline to negative was the most frequent(20 of the 31 epithelial layers). Patients with positive margin in the total and epithelial layers by FSD demonstrated a significantly worse local recurrence-free survival(RFS) compared with patients with borderline and negative margins, which revealed comparable local RFS. Patients with borderline and negative margins in the epithelial layer by PSD also revealed comparable local RFS. These results suggested that epithelial borderline might be regarded substantially as negative. When classifying the status of the epithelial layer either as negative or positive, concordance rates between FSD and PSD in the total, epithelial, and subepithelial layers were 95.0%, 93.0%, and 98.0%, respectively.CONCLUSION During intraoperative assessment of bile duct margin, borderline in the epithelial layer can be substantially regarded as negative, under which condition FSD is comparable to PSD.展开更多
文摘针对现有皮肤镜图像分割技术分割精度不高的问题,提出了一种基于多尺度信息提取和特征融合的U型网络(multi-scale information extraction and feature fusion U-shaped network,MF-UNet)模型。在U-Net的基础上,在卷积层后加入批归一化层,将原本的跳跃连接部分替换为4级特征融合模块,充分利用语义信息和位置信息,在特征提取端末尾加入多尺度空洞卷积模块和多尺度池化模块,增大感受野,利用双路拼接上采样模块进行上采样,减少图像恢复过程中的信息损失。实验表明,相较于U-Net模型,MF-UNet在平均交并比(mean intersection over union,MIoU)上提升了14.32%,在戴斯相似系数(Dice similarity coefficient,DSC)上提升了13.18%,取得了较好的结果。该研究为计算机技术辅助医生进行皮肤病诊断提供了借鉴。
基金JSPS KAKENHI(No.JP16K08695)from the Ministry of Education,Culture,Sports,Science and Technology of Japan
文摘AIM To evaluate the usefulness of frozen section diagnosis(FSD) of bile duct margins during surgery for extrahepatic cholangiocarcinoma(CCA). METHODS We retrospectively analyzed 74 consecutive patients who underwent surgery for extrahepatic CCA from 2012 to 2017, during which FSD of bile duct margins was performed. They consisted of 40 distant and 34 perihilar CCAs(45 and 55 bile duct margins, respectively). The diagnosis was classified into three categories: negative, borderline(biliary intraepithelial neoplasia-1 and 2, and indefinite for neoplasia), or positive. FSD in the epithelial layer, subepithelial layer, and total layer was compared with corresponding permanent section diagnosis(PSD) postoperatively.Then, association between FSD and local recurrence was analyzed with special reference to borderline.RESULTS Analysis of 100 duct margins revealed that concordance rate between FSD and PSD was 68.0% in the total layer, 69.0% in the epithelial layer, and 98.0% in the subepithelial layer. The extent of remaining biliary epithelium was comparable between FSD and PSD, and more than half of the margins lost > 50% of the entire epithelium, suggesting low quality of the samples. In FSD, the rate of negative margins decreased and that of borderline and positive margins increased according to the extent of the remaining epithelium. Diagnostic discordance between FSD and PSD was observed in 31 epithelial layers and two subepithelial layers. Alteration from borderline to negative was the most frequent(20 of the 31 epithelial layers). Patients with positive margin in the total and epithelial layers by FSD demonstrated a significantly worse local recurrence-free survival(RFS) compared with patients with borderline and negative margins, which revealed comparable local RFS. Patients with borderline and negative margins in the epithelial layer by PSD also revealed comparable local RFS. These results suggested that epithelial borderline might be regarded substantially as negative. When classifying the status of the epithelial layer either as negative or positive, concordance rates between FSD and PSD in the total, epithelial, and subepithelial layers were 95.0%, 93.0%, and 98.0%, respectively.CONCLUSION During intraoperative assessment of bile duct margin, borderline in the epithelial layer can be substantially regarded as negative, under which condition FSD is comparable to PSD.