深度学习DenseNet模型在颈部淋巴结定性诊断中应用价值  

Value of deep learning DenseNet model to qualitative diagnosis of cervical lymph nodes

作　　者：张惠柯 秦若曦 史大鹏 陈传亮 ZHAN GHuike;QIN Ruoxi;SHI Dapeng;CHEN Chuanliang(Department of Medical Imaging,Henan Provincial People's Hospital,Zhengzhou University People's Hospital,Zhengzhou,Henan 450003,China;Academy of Information Systems Engineering,Information Engineering University,Zhengzhou,Henan 45000l,China)

机构地区：[1]河南省人民医院医学影像科郑州大学人民医院,河南郑州450003 [2]信息工程大学信息系统工程学院,河南郑州450001

出　　处：《中华实用诊断与治疗杂志》2025年第2期164-169,共6页Journal of Chinese Practical Diagnosis and Therapy

摘　　要：目的 探讨深度学习DenseNet模型在颈部淋巴结良恶性诊断中的应用价值,为临床诊疗提供依据。方法 2017年9月—2019年5月河南省人民医院诊治淋巴结肿大患者115例,经穿刺活检或手术切除共566枚淋巴结,其中组织病理检查诊断良性359枚,恶性207枚。按照就诊时间顺序将566枚淋巴结中前486枚(良性317枚,恶性169枚)纳入训练集,后80枚(良性42枚,恶性38枚)纳入验证集。患者淋巴结穿刺活检或手术前2周内行颈部增强CT检查,获取淋巴结轴位图像,训练集采用DenseNet模型进行训练,并对验证集淋巴结进行定性诊断。采用深度学习Python模型的Scikit-learn工具包,以组织病理检查结果为金标准,计算DenseNet模型诊断验证集颈部淋巴结良恶性的灵敏度、特异度、准确率、阳性预测值、阴性预测值。绘制ROC曲线,评估DenseNet模型诊断验证集颈部淋巴结良恶性的效能。结果 566枚淋巴结中,组织病理检查诊断良性359枚,短轴直径(8.67±3.10)mm;原发病因为反应性增生129枚,结核66枚,淋巴结炎22枚,Castlman病17枚,正常淋巴结125枚。组织病理检查诊断恶性207枚,短轴直径(14.84±5.87)mm;原发病因为肺癌71枚,喉癌37枚,甲状腺癌24枚,舌癌15枚,神经内分泌癌14枚,淋巴瘤8枚,乳腺癌8枚,卵巢癌7枚,涎腺导管癌7枚,鼻咽癌5枚,食管癌4枚,尿路上皮癌3枚,前列腺癌2枚,牙龈癌1枚,下颌骨癌1枚。验证集组织病理检查诊断为良性的42枚淋巴结中,DenseNet模型准确诊断38枚,误判为恶性4枚;组织病理检查诊断为恶性的38枚淋巴结中,DenseNet模型准确诊断29枚,误判为良性9枚。以组织病理检查结果为金标准,DenseNet模型诊断验证集颈部淋巴结良恶性的灵敏度为76.3%,特异度为90.5%,准确率为83.8%,阳性预测值为87.9%,阴性预测值为80.9%。DenseNet模型诊断验证集颈部淋巴结良恶性的AUC为0.842(95%CI:0.796~0.904,P=0.030),灵敏度为76.3%,特异度为90.5%。DenseNet模�Objective To explore the application value of deep learning DenseNet model to the diagnosis of benign and malignant cervical lymph nodes,and to provide a basis for clinical diagnosis and therapy.Methods Totally 115 patients with lymphadenectasis were diagnosed and treated in Henan Provincial People's Hospital from September 2017 to May 2019,and 566 lymph nodes were performed puncture biopsy or surgical excision,among which 359 were benign and 207 were malignant by histopathological examination.According to the sequence of patient encounter time,the first 486 of 566 lymph nodes(317 benign and 169 malignant)were included in the training set,and the last 80 lymph nodes(42 benign and 38 malignant)were included in the validation set.The cervical enhanced CT examination was performed to obtain axial lymph node images within 2 weeks before puncture biopsy or surgery.DenseNet model was used for training in the training set,and qualitative diagnosis of lymph nodes in the verification set was performed.According to the histopathological examination results as the gold standard,the deep learning Python model's Scikit-learn toolkit was used to calculate the sensitivity,specificity,accuracy,positive predictive value and negative predictive value of the diagnosis of benign and malignant cervical lymph nodes by DenseNet model in the validation set.ROC curve was plotted to evaluate the efficiency of DenseNet model on diagnosing benign and malignant cervical lymph nodes in the validation set.Results Among 566 lymph nodes,359 were benign by histopathological examination,with the short axis diameter of(8.67±3.10)mm.The primary diseases were reactive hyperplasia in 129 nodes,tuberculosis in 66,lymphadenitis in 22,Castlman disease in 17,and normal lymph nodes in 125.Totally 207 were malignant by histopathological examination,with the short axis diameter of(14.84±5.87)mm.The primary diseases were lung cancer in 71 nodes,laryngeal cancer in 37,thyroid cancer in 24,tongue cancer in 15,neuroendocrine cancer in 14,lymphoma in 8,breast c

关 键 词：颈部淋巴结 深度学习 DenseNet模型 CT检查 

分 类 号：R551.2[医药卫生—血液循环系统疾病] TP18[医药卫生—内科学]