基于多模态图像构建CNN-ViT模型在弥漫性大B细胞淋巴瘤骨髓受累诊断中的应用  

Application of CNN-ViT Fusion Model Based on Multimodal Images in the Diagnosis of Bone Marrow Involvement in Diffuse Large B-Cell Lymphoma

作　　者：李兰兰 周颖 林禹 尤梦翔 林美福[2,3] 陈文新[2,3] LI Lanlan;ZHOU Ying;LIN Yu;YOU Mengxiang;LIN Meifu;CHEN Wenxin(不详;Shengli Clinical Medical College of Fujian Medical University,Fuzhou 350001,China;Department of Nuclear Medicine,FujianProvincial Hospital,Fuzhou 350001,China)

机构地区：[1]福州大学物理与信息工程学院福建省媒体信息智能处理与无线传输重点实验室,福建闽侯350100 [2]福建医科大学省立临床医学院,福建福州350001 [3]福建省立医院核医学科,福建福州350001 [4]宁德师范学院附属宁德市医院,福建宁德352000

出　　处：《中国医学影像学杂志》2023年第4期390-394,共5页Chinese Journal of Medical Imaging

摘　　要：目的设计一种融合多模态图像深度学习模型CNN-ViT,诊断弥漫性大B细胞淋巴瘤(DLBCL)骨髓受累。资料与方法回顾性收集2012年11月—2022年6月福建省立医院经病理证实的DLBCL 78例,其中无骨髓受累46例,有骨髓受累32例,所有患者在化疗前均行全身18F-FDG PET/CT检查、骨髓穿刺细胞涂片和(或)骨髓活检。选取骨盆区域PET及CT图像共9828张。将上述数据按7∶1∶2随机分为训练集6858张、验证集982张和测试集1988张。结合传统的卷积神经网络(CNN)和Vision-Transformer(ViT)模型设计CNN-ViT模型,分别提取PET和CT图像特征,预测骨髓受累情况。使用测试集的混淆矩阵和损失函数的变化、准确度、敏感度、特异度和F1_score评价模型的性能。结果CNN-ViT模型诊断DLBCL骨髓受累的准确度、特异度、敏感度和F1_score分别为0.988、0.971、0.997、0.987。结论CNN-ViT模型可以准确评估DLBCL骨髓受累情况。Purpose To design a fused multimodal image deep learning model(convolutional neural networks with Vision-Tranformer,CNN-ViT)to diagnose bone marrow involvement in patients with diffuse large B-cell lymphoma(DLBCL).Material and Methods A total of 78 patients with pathologically confirmed DLBCL from November 2012 to June 2022 in Fujian Provincial Hospital were collected retrospectively,dividing into without bone marrow involvement group(n=46)and with bone marrow involvement group(n=32).All the patients underwent a whole-body 18F-FDG-PET/CT scan,bone marrow smear and/or biopsy before chemotherapy.PET and CT images of the pelvic region were selected(9828 images in total).All image data were divided randomly into a training set(70%,n=6858 sheets),validation set(10%,n=982 sheets),and test set(20%,n=1988 sheets)at a ratio of 7∶1∶2.A CNN-ViT learning model was designed via combination between the traditional CNN and ViT models to extract the features of PET and CT images and to evaluate the situation of bone marrow involvement.The confusion matrix,loss function,accuracy,specificity,sensitivity and F1_score were used to evaluate the performance of the classification model.Results The performance of the fusion multimodal earning model such as CNN-ViT implemented in the diagnosis of DLBCL bone marrow involvement was good,with the accuracy of 0.988,the specificity of 0.971,the sensitivity of 0.997,and the F1_score was 0.987,respectively.Conclusion A deep learning model fused with multimodal images,such as CNN-ViT,shows good performance in accurately diagnosing bone marrow involvement in patients with DLBCL.

关 键 词：淋巴瘤 B细胞 正电子发射断层显像术 体层摄影术 X线计算机 骨髓 神经网络 骨盆 

分 类 号：R733.4[医药卫生—肿瘤] R445.6[医药卫生—临床医学]