This work aimed to explore chest computed tomography (CT) image segmentation of patients with large-cell neuroendocrine carcinoma (LCNEC) based on deep learning, as well as the clinical manifestations and imaging and pathological features of LCNEC patients. Clinical data of 40 patients with LCNEC confirmed by pathological examination in the X Hospital from December 2015 to December 2017 were retrospectively selected. CT image data were segmented by TJ-1 model modified full convolutional neural network (FCNN) model. The accuracy and training time of TJ-1 FCNN model and classic deep learning segmentation network model AlexNet model were compared in terms of image segmentation. According to the image segmentation results by TJ-1 FCNN model, chest CT images of LCNEC patients, were reviewed, and the clinical manifestations, as well as the imaging and pathological features of the patients were reviewed, sorted, and summarized. The results showed that the image segmentation accuracy of TJ-1 network model (99.38%) was higher than that of AlexNet model. The iteration training time of TJ-1 network model for 30 times was 45 min, lower than that of AlexNet model (82 min). LCNEC was more likely to be found in elderly male with a long history of smoking. The clinical symptoms were cough, sputum, sputum blood, and chest pain with no significant specificity. CT imaging showed that peripheral mass was the most common manifestation (67.5%), both lungs were visible, the upper lobe was more likely with lesion (60%), the edge of the lesion was clear or smooth (57.5%), which was lobulated (70%). Under the light microscope, tumor cells were characterized by large volume, low nucleocytoplasmic ratio, high mitosis, and large area necrosis. The positive rates of immunohistochemical neuroendocrine markers were CD56 (62.5%), CgA (50%), and Syn (85%), among which Syn was with the highest positive rate. To sum up, LCNEC lacked clinical and radiological specificity manifestations, while chest CT image segmentation based on TJ-1 FCNN model can quickly mark the location of the lesion, providing technical support for the diagnosis and evaluation of LCNEC clinically.

这项工作旨在探索基于深度学习的大细胞神经内分泌癌（LCNEC）患者的胸部计算机断层扫描（CT）图像分割，以及LCNEC患者的临床表现，影像学和病理学特征。回顾性选择2015年12月至2017年12月在X医院经病理检查确诊的40例LCNEC患者的临床资料。通过TJ-1模型修正的全卷积神经网络（FCNN）模型对CT图像数据进行分割。从图像分割的角度比较了TJ-1 FCNN模型和经典深度学习分割网络模型AlexNet模型的准确性和训练时间。根据TJ-1 FCNN模型的图像分割结果，回顾了LCNEC患者的胸部CT图像，并对其临床表现进行了总结，以及对患者的影像学和病理学特征进行了回顾，分类和总结。结果表明，TJ-1网络模型的图像分割精度为99.38％，高于AlexNet模型。TJ-1网络模型的30次迭代训练时间为45分钟，低于AlexNet模型的82分钟。LCNEC更有可能在具有悠久吸烟史的老年男性中发现。临床症状为咳嗽，痰，痰血和胸痛，无明显特异性。CT扫描显示，周围肿块是最常见的表现（67.5％），双肺均可见，上叶更易发生病变（60％），病变边缘清晰或光滑（57.5％），小叶（70％）。在光学显微镜下，肿瘤细胞的特征是体积大，核质比低，有丝分裂高，坏死面积大。免疫组化神经内分泌标志物的阳性率为CD56（62.5％），CgA（50％）和Syn（85％），其中Syn阳性率最高。综上所述，LCNEC缺乏临床和放射学特异性表现，而基于TJ-1 FCNN模型的胸部CT图像分割可以快速标记病变部位，为临床诊断和评价LCNEC提供技术支持。