基于深度学习的溃疡性结肠炎分级方法  

Ulcerative Colitis Grading Method Based on Deep Learning

作　　者：刘燕红 张俊华[1] 缪佳蓉[2] 罗旭东 LIU Yan-hong;ZHANG Jun-hua;MIAO Jia-rong;LUO Xu-dong(School of Information Science and Engineering,Yunnan University,Kunming Yunnan 650500,China;First Affiliated Hospital,Kunming Medical University,Kunming Yunnan 650032,China)

机构地区：[1]云南大学信息学院,云南昆明650500 [2]昆明医科大学第一附属医院,云南昆明650032

出　　处：《计算机仿真》2024年第12期284-290,共7页Computer Simulation

基　　金：国家自然科学基金资助项目(62063034);云南省高层次卫生技术人才培养经费资助(H-2018040)。

摘　　要：针对溃疡性结肠炎(Ulcerative Colitis, UC)分级诊断工作量大、主观性强的问题,提出了一种基于DenseNet的自动分级方法。引入坐标注意机制和级联型特征金字塔网络改进DenseNet,以增强多层次特征提取能力,并采用焦点损失函数解决多分类样本不均衡问题,最后利用测试时增强技术提升模型整体鲁棒性。相比原网络,提出的方法在两个数据集上评估UC是否内镜缓解的准确率提高了1.0%和0.8%;评估内镜下病变严重程度的准确率提高了1.2%和2.4%。上述方法提高了UC疾病评估的精度,对临床辅助诊断有一定的参考价值和意义。The grading diagnosis of ulcerative colitis(UC)is laborious and subjective.To address the problem,an automatic grading method based on DenseNet was proposed.The multi-level feature extraction capability of DenseNet was enhanced by introducing a coordinate attention mechanism and a cascaded feature pyramid network.Besides,the focal loss function was used to solve the problem of imbalanced multi-class samples.Finally,test-time augmentation technology was used to improve the overall robustness of the model.Compared with the origin model,the proposed method improved the accuracy of assessing whether UC was in endoscopic remission by 1%and 0.8%,and the accuracy of assessing the severity of endoscopic lesions was improved by 1.2%and 2.4%,respectively on two datasets.The method improves the accuracy of UC disease assessment,having certain reference values and significance for clinical auxiliary diagnosis.

关 键 词：溃疡性结肠炎 卷积神经网络 坐标注意机制 级联型特征金字塔网络 焦点损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]