Deep Learning Computer-aided Diagnostic Model for Non-small Cell Lung Cancer Based on Convolutional Neural Network and Attention Mechanism
-
摘要:
目的：构建基于卷积神经网络（CNN）和注意力机制的改进CNN模型（ISANET），评估模型的性能，并与传统CNN模型进行比较。方法：收集经手术病理证实的肺鳞癌（LUSC）或肺腺癌（LUAD）患者60例以及肺部正常患者30例的肺部CT平扫或增强图像共619张，组成DatasetA；收集公共数据集图像共737张，组成DatasetB。两数据集均按6∶4将随机分为训练集和测试集。构建ISANET模型并进行训练和验证，然后记录查准率、召回率，并计算出F1分数，用以评价ISANAT模型的效能。最后，将ISANET模型与传统CNN模型AlexNet，VGG 16，Inception V3，Mobilenet V2，ResNet 18进行对比，绘制P-R（P-R）曲线，计算出P-R曲线下面积，并评估不同模型对肺鳞癌和肺腺癌的鉴别效能。结果：相较于传统CNN模型，ISANET模型对非小细胞肺癌分类的准确度明显提高，在DatasetA和DatasetB中分别为99.6%和95.2%。结论：ISANET模型较好地实现了对肺鳞癌和肺腺癌的无创预测，提高了肺鳞癌和肺腺癌CT影像鉴别的准确度，能够帮助诊断医师对非小细胞肺癌进行快速准确的分类。
Abstract:Objective: To construct an improved deep learning computer-aided diagnosis model based on convolutional neural network (CNN) and Attention Mechanism proposed as Inception Spatial and Channel Attention Network (ISANET) and evaluate the model's performance, comparing it with the traditional CNN model. Methods: A total of 619 lung CT images of 60 patients with lung squamous cell carcinoma or lung adenocarcinoma confirmed by surgical pathology and 30 patients with normal lungs were collected retrospectively to form Dataset A; a total of 737 public dataset images were collected to form Dataset B. The two datasets were randomly divided into training and test sets at a 6:4 ratio. Construct the ISANET model and conduct training and verification, then record the precision ratio and recall ratio to calculate the F1 score and evaluate the performance of the ISANAT model. Finally, the ISANET model was compared with the traditional CNN models such as AlexNet, VGG16, InceptionV3, MobilenetV2, and ResNet18 by drawing the Precision-Recall (P-R) curve and calculating the area under the P-R curve to evaluate the classification performance of different models for LUSC and LUAD. Results: Compared with the traditional CNN model, the accuracy of the ISANET model for non-small cell lung cancer classification improved significantly, reaching 99.6% and 95.2% in Dataset A and Dataset B, respectively. Conclusions: The ISANET model provides better non-invasive prediction of LUSC and LUAD, improves the accuracy of CT imaging identification of lung squamous cell carcinoma and lung adenocarcinoma, and can help diagnosticians quickly and accurately classify non-small cell lung cancer.
-
Keywords:
- non-small cell lung cancer /
- CNN /
- attention mechanisms /
- computer-aided diagnosis
-
1. 病例资料
患者女，42岁，因外院体检发现胸部占位10余天，为求进一步手术治疗就诊。患者既往体健，无明显呼吸困难、胸痛等不适。查体：脊柱侧弯，肝浊音界下移，两肺无明显干湿啰音，两肺呼吸音不对称，右中下肺野呼吸音消失，叩诊呈实音。实验室检查：血常规、肝肾功能、肿瘤标记物、大小便常规未见明显异常。
胸部CT平扫及增强检查：右后纵隔巨大稍低密度肿块，平扫密度低于邻近肌肉，CT值约26 HU，边界清楚，内见数枚点状钙化（图1（a））；肿块内部近边缘处可见结节状脂肪密度影，CT值约-50 HU，脊柱明显侧弯（图1（b））；增强扫描动脉期肿块强化不明显，CT值约28 HU，椎旁可见供血动脉伸入肿块内部（图1（c））；静脉期最大密度投影（maximum intensity projection，MIP）显示肿瘤供血动脉由胸主动脉发出（图1（d））；静脉期冠状面重建肿块强化仍不明显，CT值约32 HU，肿块沿人体长轴生长，最大截面约17.2 cm×13.3 cm，边界清楚，肺和膈肌受推挤，肝脏下移（图1（e））；延迟期肿块轻度不均匀强化，CT值约37 HU；肿块呈嵌入式生长，向胸椎和主动脉前方延伸，与主动脉分界尚清（图1（f））；术前CT诊断：①节细胞神经瘤（ganglioneuroma，GN）；②黏液样脂肪肉瘤。
![]()
图 1 纵隔巨大GN的CT图像（a）、（b）轴位CT平扫，右后纵隔稍低密度肿块，白箭示肿块内点状钙化（a）、边缘结节状脂肪成分（b）；（c）增强扫描动脉期，肿块强化不明显，白箭示供血动脉；（d）静脉期MIP图，白箭示肿瘤供血动脉自胸主动脉发出；（e）静脉期冠状面重建，显示肿块最大截面大小（蓝黄线）；（f）延迟期，肿块轻度强化，白箭示嵌入式生长。Figure 1. CT images of a large GN in the mediastinum手术治疗：全麻下行胸腔镜辅助纵隔肿瘤切除术+胸膜粘连松解术。麻醉满意后取左侧卧位，胸下垫枕，常规消毒铺巾，取右侧第6肋后外侧切口入路，长径20 cm，逐层切开皮肤、皮下组织、肌肉后，沿肋间切开肋间肌肉后，进胸。右侧第8肋间长约1 cm切口，作为观察孔。术中探查见：胸膜少许粘连，胸腔无积液。脊柱侧弯明显，肿瘤位于右后纵隔，大小约20 cm×18 cm，包膜完整，质稍硬，活动度不佳，部分深入脊柱后方。部分右下肺与肿瘤粘连紧密。分离胸膜粘连，沿肿瘤边缘切开包膜后，钝性+锐性逐步游离肿瘤，见肿瘤后壁深入脊柱后方，界限不清，无法暴露，遂沿脊柱表面电刀完整切除肿瘤。彻底止血后，查无活动性出血，清点纱布器械无误，于右第8肋间置胸引管一根，逐层关胸。术中出血不多，未输血。标本家属过目后送病理检查。
病理及免疫组化：肿物剖面灰白、灰黄色、质软，大小17 cm×16 cm×8.5 cm。光镜下可见成熟的神经节细胞和Schwann细胞（图2（a））；节细胞和梭形神经组织免疫标记S100呈阳性（2（b））。免疫组化：CD117（-），CD34（脉管+），CD56（+），CgA（+），Desmin（-），NSE（+），S100（+），SMA（-），Syn（+），Ki-67（个别细胞+）。病理诊断：（纵隔）GN。
患者术后出现一定量右侧胸腔积液，经治疗好转后出院，出院后1个月于我院胸外科随诊，复查胸部CT，无明显肿瘤残留，右侧胸腔积液基本吸收，3个月后继续随诊一次，患者恢复良好，无特殊不适。
2. 讨论
GN由神经节细胞、Schwann细胞和纤维组织构成[1]，常无明显临床症状，当肿块巨大时，可产生压迫症状，表现为呼吸困难、咳嗽、胸闷、腹胀等。GN起源于交感神经链，占神经源性肿瘤的2％～3％[2]，可发生于任何年龄，20岁前发病者约占60％，女性发病率稍高于男性。大部分GN不分泌生物活性物，少部分GN可分泌儿茶酚胺、血管活性肠肽等物质，患者从而出现相应临床症状，如高血压、腹泻等[3]。少数GN可合并神经纤维瘤病，有观点认为，GN可能是神经纤维瘤病的局部表现之一，也有学者持不同观点，关于GN与神经纤维瘤病两者间是否存在某种内在联系，还有待进一步深入研究。
根据目前文献回顾，本例是迄今报道的成人最大纵隔GN之一[4-5]，GN常在影像学检查时偶然发现，纵隔GN常表现为边界清楚的软组织肿块，多伴有包膜。显微镜下，GN常表现为神经纤维组织内散在分布成熟神经节细胞；S100、Vim、NSE、髓磷脂碱性蛋白等免疫组织化学染色可为阳性[6]。本例肿块巨大，存在恶变、持续生长的风险，考虑到活检结果不会影响治疗策略，且穿刺活检可伴发潜在的并发症，如气胸、胸膜和活检道肿瘤种植等，故本例采取直接手术切除进行治疗。
GN的CT表现与其病理及生物学特点相关，病理上GN存在包膜，CT对包膜的直接显示率不足，但肿瘤边界清楚，可间接反映其生长受包膜限制。GN含大量黏液基质，水分含量高，CT平扫肿瘤密度常低于邻近肌肉。GN的质地较软，可呈嵌入式或塑形样紧贴椎体生长，也可侵入椎管[7]、椎前，呈“伪足样”包绕邻近大血管，但侵蚀骨质者极为罕见[8]。此外，GN多沿交感神经链即人体长轴生长，头尾径大于人体短轴径[9]，Ozawa等[10]研究发现GN头尾径与人体短轴径比值为1.50，明显大于神经鞘瘤（0.95～1.10）。GN有时成分复杂，可伴砂砾样钙化，少数肿块边缘还可伴有少量脂肪组织，本例亦可观察到此类表现。GN内出现脂肪组织原因可能是肿瘤生长、包埋邻近区域脂肪所造成的。也有学者认为，GN内的脂肪是由于肿瘤长期生长，组织退变所造成的[11]。GN为乏血供肿瘤，多呈轻度延迟强化，有时可见肋间供血动脉，肿瘤内含有较粗大、互相交错排列的细胞成分和神经纤维的区域通常强化程度较低，大量黏液基质可引起造影剂滞留，出现延迟强化[12]。Yorita等[13]报道的一例纵隔巨大GN，平扫CT值27 HU，增强扫描延迟期CT值仅上升7 HU，与本例强化方式类似。
GN的CT表现有时不典型，此外，鉴于纵隔肿瘤的广泛鉴别诊断，GN术前CT诊断有时仍具挑战性[14]，其主要鉴别诊断包括：①神经母细胞瘤和节细胞神经母细胞瘤，发病年龄更小，常发生于婴幼儿、儿童，儿茶酚胺常上升，病变具一定侵袭性，肿块内部可合并出血、坏死，钙化更常见，常为粗大、不规则形钙化；增强明显强化，可伴血行转移[15]。②神经鞘瘤和神经纤维瘤，神经鞘瘤多呈卵圆形、外形规则，神经鞘瘤中Antoni A区细胞丰富、明显强化；Antoni B区细胞排列疏松，易囊变，无明显强化[16]。神经纤维瘤常伴随神经纤维瘤病，肿块呈结节状，边界欠清，增强中等强化。无论是神经鞘瘤还是神经纤维瘤大多无延迟强化的特点，且多沿人体短轴生长。③黏液样脂肪肉瘤，主要发生于四肢深部和腹膜后，张慧红等[17]总结的18个病例中发生于纵隔者仅1例，黏液样脂肪肉瘤密度和强化特征可与GN类似，但其内部可有分隔、钙化少见，脂肪成分占比多＞10％。
综上，GN典型CT表现为边界清楚的软组织肿块，可伴钙化和脂肪、呈嵌入式沿人体长轴生长，轻度延迟强化，CT术前诊断价值较高。
-
![]()
图 5 ISANET模型在不同数据集中的准确率和损失曲线
Figure 5. Accuracy and loss curves of ISANET on different datasets
表 1 不同模型在DatasetA中的查准率、召回率和F1值比较情况
Table 1 Comparison of precision, sensitivity, and F1 score of different models in dataset A
算法名称 腺癌 鳞癌 未见明显异常 查准率 召回率 F1值 查准率 召回率 F1值 查准率 召回率 F1值 ISANET 0.898 0.925 0.906 0.946 0.914 0.928 0.959 0.985 0.968 AlexNet 0.870 0.852 0.854 0.863 0.895 0.872 0.994 0.999 0.996 VGG16 0.798 0.789 0.761 0.764 0.857 0.771 0.995 0.992 0.994 InceptionV3 0.974 0.807 0.873 0.754 0.941 0.824 0.961 0.954 0.951 MobileNetV2 0.640 0.902 0.733 0.933 0.711 0.799 0.959 0.963 0.953 ResNet18 0.919 0.929 0.919 0.927 0.905 0.911 0.960 0.992 0.969 
下载: 导出CSV
表 2 不同模型在DatasetB中的查准率、召回率和F1值比较情况
Table 2 Comparison of precision, sensitivity, and F1 score of different models in dataset B
算法名称 腺癌 鳞癌 未见明显异常 查准率 召回率 F1值 查准率 召回率 F1值 查准率 召回率 F1值 ISANET 0.911 0.824 0.862 0.754 0.882 0.808 0.937 0.942 0.936 AlexNet 0.846 0.740 0.779 0.575 0.816 0.646 0.992 0.895 0.941 VGG16 0.814 0.720 0.717 0.507 0.668 0.531 0.964 0.949 0.952 InceptionV3 0.883 0.835 0.849 0.758 0.824 0.777 0.876 0.938 0.882 MobileNetV2 0.896 0.748 0.805 0.601 0.858 0.670 0.939 0.904 0.910 ResNet18 0.831 0.866 0.842 0.792 0.785 0.788 0.989 0.874 0.926
下载: 导出CSV
表 3 各模型在不同数据集中的准确度
Table 3 Accuracy of each model in different datasets
算法名称 Dataset A Dataset B ISANET 0.996 0.952 AlexNet 0.951 0.854 VGG16 0.939 0.901 InceptionV3 0.980 0.946 MobileNetV2 0.964 0.932 ResNet18 0.992 0.939
下载: 导出CSV
表 4 消融实验结果
Table 4 The results of ablation experiments
组别 Dataset A Dataset B A组 0.806 0.745 B组 0.810 0.735 C组 0.854 0.796 D组 0.802 0.721
下载: 导出CSV
-
[1] HERBST R S, MORGENSZTERN D, BOSHOFF C. The biology and management of non-small cell lung cancer[J]. Nature, 2018, 553(7689): 446-454. DOI: 10.1038/nature25183.
[2] 中华医学会肿瘤学分会, 中华医学会杂志社. 中华医学会肺癌临床诊疗指南(2022版)[J]. 中华肿瘤杂志, 2022, 44(6): 457-490. DOI: 10.3760/cma.j.cn112152-20220413-00255. ONCOLOGY SOCIETY OF CHINESE MEDICAL ASSOCIATION, CHINESE MEDICAL ASSOCIATION PUBLISHING HOUSE. Chinese Medical Association guideline for clinical diagnosis and treatment of lung cancer (2022 edition)[J]. Chinese Journal of Oncology, 2022, 44(6): 457-490. DOI: 10.3760/cma.j.cn112152-20220413-00255.
[3] CARRILLO-PEREZ F, MORALES J C, CASTILLO-SECILLA D, et al. Non-small-cell lung cancer classification via RNA-Seq and histology imaging probability fusion[J]. BMC Bioinformatics, 2021, 22(1): 454. DOI: 10.1186/s12859-021-04376-1.
[4] WOODARD G A, JONES K D, JABLONS D M. Lung cancer staging and prognosis[J]. Cancer Treatment and Research, 2016, 170: 47-75. DOI: 10.1007/978-3-319-40389-2_3.
[5] CAO B, WANG P, GU L, et al. Use of four genes in exosomes as biomarkers for the identification of lung adenocarcinoma and lung squamous cell carcinoma[J]. Oncology Letters, 2021, 21(4): 249. DOI: 10.3892/ol.2021.12510.
[6] KIM T-H, WOO S, YOON S H, et al. CT Characteristics of non–small cell lung cancer with anaplastic lymphoma kinase rearrangement: A systematic review and meta-analysis[J]. American Journal of Roentgenology, 2019, 213(5): 1059-1072. DOI: 10.2214/AJR.19.21485.
[7] 宋若晨, 褚相乐, 黄勇华, 等. 基于颅脑T1WI对比增强图像构建卷积神经网络模型鉴别肺癌与乳腺癌脑转移[J]. 中国医学影像技术, 2023, 39(7): 982-986. SONG R C, CHU X L, HUANG Y H, et al. Convolutional neural network model based on contrast-enhanced cranial T1WI for differentiating brain metastases from lung cancer or breast cancer[J]. Chinese Journal of Medical Imaging Technology, 2023, 39(7): 982-986.
[8] LITJENS G, KOOI T, BEJNORDI B E, et al. A survey on deep learning in medical image analysis[J]. Medical Image Analysis, 2017, 42: 60-88. DOI: 10.1016/j.media.2017.07.005.
[9] YAMASHITA R, NISHIO M, DO R K G, et al. Convolutional neural networks: An overview and application in radiology[J]. Insights into Imaging, 2018, 9(4): 611-629. DOI: 10.1007/s13244-018-0639-9.
[10] RUSSAKOVSKY O, DENG J, SU H, et al. ImageNet large scale visual recognition challenge[J]. International Journal of Computer Vision, 2015, 115(3): 211-252. DOI: 10.1007/s11263-015-0816-y.
[11] SZEGEDY C, VANHOUCKE V, IOFFE S, et al. Rethinking the inception architecture for computer vision[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV, USA: IEEE, 2016: 2818-2826.
[12] COUDRAY N, OCAMPO P S, SAKELLAROPOULOS T, et al. Classification and mutation prediction from non–small cell lung cancer histopathology images using deep learning[J]. Nature Medicine, 2018, 24(10): 1559-1567. DOI: 10.1038/s41591-018-0177-5.
[13] JIANG X, SHEN H. Classification of lung tissue with cystic fibrosis lung disease via deep convolutional neural networks[C]//Proceedings of the 2nd International Symposium on Image Computing and Digital Medicine. Chengdu China: ACM, 2018: 113-116.
[14] MOITRA D, KR. MANDAL R. Classification of non-small cell lung cancer using one-dimensional convolutional neural network[J]. Expert Systems with Applications, 2020, 159: 113564. DOI: 10.1016/j.eswa.2020.113564.
[15] 唐秉航, 王艳芳, 马力, 等. 基于混合注意力机制的肺结节假阳性降低[J]. CT理论与应用研究, 2022, 31(1): 63-72. DOI: 10.15953/j.ctta.2021.002. TANG B H, WANG Y F, MA L, et al. False positive reduction of pulmonary nodules based on mixed attentional raechanism[J]. CT Theory and Applications, 2022, 31(1): 63-72. DOI: 10.15953/j.ctta.2021.002.
[16] TERAMOTO A, TSUKAMOTO T, KIRIYAMA Y, et al. Automated classification of lung cancer types from cytological images using deep convolutional neural networks[J]. BioMed Research International, 2017, 2017: 1-6.
[17] TERAMOTO A, YAMADA A, KIRIYAMA Y, et al. Automated classification of benign and malignant cells from lung cytological images using deep convolutional neural network[J]. Informatics in Medicine Unlocked, 2019, 16: 100205. DOI: 10.1016/j.imu.2019.100205.
[18] 樊雪林, 文昱齐, 乔志伟. 基于Transformer增强型U-net的CT图像稀疏重建与伪影抑制[J]. CT理论与应用研究(中英文), 2024, 33(1): 1-12. DOI: 10.15953/j.ctta.2023.183. FAN X L, WEN Y Q, QIAO Z W. Sparse Reconstruction of Computed Tomography Images with Transformer Enhanced U-net[J]. CT Theory and Applications, 2024, 33(1): 1-12. DOI: 10.15953/j.ctta.2023.183.
[19] KHAN M A, RAJINIKANTH V, SATAPATHY S C, et al. VGG19 network assisted joint segmentation and classification of lung nodules in CT images[J]. Diagnostics, 2021, 11(12): 2208. DOI: 10.3390/diagnostics11122208.
[20] 王媛媛, 周涛, 陆惠玲, 等. 基于集成卷积神经网络的肺部肿瘤计算机辅助诊断模型[J]. 生物医学工程学杂志, 2017, 34(4): 543-551. DOI: 10.7507/1001-5515.201607003. WANG Y Y, ZHOU T, LU H L, et al. Computer aided diagnosis model for lung tumor based on ensemble convolutional neural network[J]. Journal of Biomedical Engineering, 2017, 34(4): 543-551. DOI: 10.7507/1001-5515.201607003.
[21] AKSELROD-BALLIN A, KARLINSKY L, ALPERT S, et al. A Region Based Convolutional Network for Tumor Detection and Classification in Breast Mammography[M]//Carneiro G, Mateus D, Peter L, et al. Deep Learning and Data Labeling for Medical Applications: Vol. 10008. Cham: Springer International Publishing, 2016: 197-205.
[22] HELWAN A, EL-FAKHRI G, SASANI H, et al. Deep networks in identifying CT brain hemorrhage[J]. Journal of Intelligent & Fuzzy Systems, 2018, 35(2): 2215-2228.
[23] ESTEVA A, KUPREL B, NOVOA R A, et al. Dermatologist-level classification of skin cancer with deep neural networks[J]. Nature, 2017, 542(7639): 115-118. DOI: 10.1038/nature21056.
[24] HARANGI B. Skin lesion classification with ensembles of deep convolutional neural networks[J]. Journal of Biomedical Informatics, 2018, 86: 25-32. DOI: 10.1016/j.jbi.2018.08.006.
[25] ALOM M Z, YAKOPCIC C, NASRIN MST S, et al. Breast Cancer Classification from Histopathological Images with Inception Recurrent Residual Convolutional Neural Network[J]. Journal of Digital Imaging, 2019, 32(4): 605-617. DOI: 10.1007/s10278-019-00182-7.
[26] SANDLER M, HOWARD A, ZHU M, et al. MobileNetV2: Inverted residuals and linear bottlenecks[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, UT: IEEE, 2018: 4510-4520.
[27] SZEGEDY C, LIU W, JIA Y Q, et al. Going deeper with convolutions[C]//2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Boston, MA, USA: IEEE, 2015: 1-9.
[28] IOFFE S, SZEGEDY C. Batch normalization: accelerating deep network training by reducing internal covariate shift[M]. arXiv, 2015. http://arxiv.org/abs/1502.03167, 2015-03-02.
[29] 王坤, 王冠民, 路志凯. 动态增强CT对周围型肺癌的诊断价值[J]. 肿瘤研究与临床, 2021, 33(4): 304-306. DOI: 10.3760/cma.j.cn115355-20201127-00667. [30] 薛彩强, 杜晓灏, 金龙, 等. 基于MRI深度学习模型预测WHOⅡ、Ⅲ级胶质瘤MGMT启动子甲基化状态[J]. 中华放射学杂志, 2021, 55(7): 734-738. XUE C K, DU X H, JIN L, et al. Prediction of methylation status of MGMT promoter in WHO grade Ⅱ, Ⅲ glioma based on MRI deep learning model[J]. Chinese Journal of Radiology, 2021, 55(7): 734-738.
[31] 孟宇, 马之骋, 阮敬儒, 等. 基于Faster R卷积神经网络构建胸部X线片异物智能检测模型的可行性研究[J]. 中华放射学杂志, 2022, 56(12): 1359-1364. MENG Y, MA Z C, RUAN J R, et al. Feasibility of constructing the intelligent detection model for foreign bodies on chest X-ray based on Faster R-convolutionalneural network[J]. Chinese Journal of Radiology, 2022, 56(12): 1359-1364.
计量
- 文章访问数: 84
- HTML全文浏览量: 8
- PDF下载量: 19
