| Background With the increasing use of low-dose spiral CT for lung cancer screening and increased awareness of physical examination,more and more patients are clinically diagnosed with lung nodules.Subsolid nodule is a kind of pulmonary nodule,they have a large proportion of malignancy,which are mainly biologically inert and have difficulty obtaining histopathology,synchronous and heterochronic multiple primary nodules are not rare.Clinical treatment often needs to consider stereotactic body radiotherapy(SBRT),but the clinical pain point is that it is often necessary to make a therapeutic decision whether to accept SBRT or not according to the clinical diagnosis.Therefore,it is necessary to carry out the research on the clinical diagnosis level of radiologists and radiation oncologists of subsolid nodules and how to improve the clinical diagnosis level according to objective indicators.This study will evaluate the clinician's qualitative diagnosis based on experience,display of one or more quantitative indexes of tumor based on chest CT and the establishment of quantitative models based on convolutional neural network(CNN)to identify potential objective indicators of tumors to identify benign and malignant subsolid nodules as the starting point.Then on the chest CT images,a CNN model based on deep learning is used to identify and analyze the biological characteristics of the malignant degree of subsolid malignant nodules,to compare and explore whether combining clinicians and some objective indicators and CNN modeling can improve the identification of malignant and malignant degrees of subsolid nodules,thereby better helping Clinical decision.Materials and methods Retrospectively collected clinical data of lung subsolid nodules undergoing surgery at Shanghai Chest Hospital from 2012 to 2018.Patients with chest CT images before operation,general clinical features and postoperative pathology were included in the group.Diagnosis is divided into three categories: 1)Clinical diagnosis on a 10-point scale: CT images of each subsolid nodule were carefully read and analyzed independently in random order by three senior chest radiation oncologists and one senior chest radiologist(score the nodule on a ten-point scale system,the higher the score,the more likely the doctor thinks the nodule is malignant).2)Quantitative diagnosis based on quantitative features of subsolid nodules in chest CT images:randomly divided the patients into training set and validation set,each patient's chest CT image was collected and the region of interest was contoured.Using the MIM graphics workstation,multiple quantitative features describing the spatial distribution of CT inside the nodule can be obtained,including nodule size and density and these quantitative features were used to establish a benign and malignant prediction model.The CT quantitative radiomic model was trained on the training set,and the efficacy of the model in identifying the benign and malignant pulmonary subsolid nodules was verified on the validation set.3)Convolutional neural network(CNN)diagnosis: Train the 3D CNN model with the training set,the network structure is modified by Dense Net and the CT images and segmentation mask images of different windows were integrated.After the training,the effectiveness of the model in identifying the benign and malignant pulmonary subsolid nodules was verified on the validation set.All the criteria of classification come from pathological diagnosis after surgery.The basic method of establishing all quantitative diagnostic classification models is to randomly divide patients into training set and validation set,collect CT images of each patient's chest and contour the region of interest,train CT quantitative radiomic model through training set,and validate the effectiveness of the model in the validation set to identify the benign and malignant pulmonary subsolid nodules.Considering the possible complementarity between different diagnoses,it is clear whether the recognition efficiency of CNN model can be further improved by combining the CNN basic model with other models.Results A total of 2614 patients who received pulmonary subsolid nodule surgery were collected,including preoperative chest CT images,clinical and postoperative pathological data(1791 cases of malignant nodules,823 cases of benign nodules).According to the general clinical characteristics of the patients,they were randomly divided into training set(2092 cases,80%)and validation set(522 cases,20%).1.The AUC values of 3 radiation oncologists identifying benign and malignant subsolid nodules were 0.815,0.695,and 0.667,respectively.The AUC value of 1radiologist identifying benign and malignant subsolid nodules was 0.877.The diagnostic efficacy of radiologist was better than that of all radiation oncologists(Oncologist1: p = 0.009,Oncologist2/3: P <0.001),and all the doctor's diagnosis efficiency was significantly better than that of the nodule maximum diameter model(AUC = 0.595,P <0.05).2.The comprehensive model of CT quantitative radiomics established by the training set has an AUC value of 0.786 for identifying benign and malignant subsolid nodules and achieved better recognition performance than the six independent features,whose respective AUCs were: the integral total mean ratio change was 0.701,the sphere value was 0.653,the volume value was 0.693,the voxel count was0.623,the integral total value was 0.695 and the total HU value was 0.628.CT quantitative radiomics combined doctor diagnosis model has an AUC value of 0.916,the sensitivity and specificity of which was 0.859 and 0.830 at best decision point respectively.3.The 3D CNN model established on the training set identifies the benign and malignant subsolid nodules with an AUC value of 0.925(95% CI: 0.899-0.950),and at the best decision point in the validation set the sensitivity was 0.854(95% CI: 0.818-0.890),the specificity was 0.876(95% CI: 0.824-0.928).4.The AUC of CNN model fusion CT quantitative radiomic model was 0.931,sensitivity was 0.873 and specificity was 0.863.The CNN model integrated doctor diagnosis model has achieved the best recognition performance,its AUC was 0.961(95%CI: 0.944-0.979),the sensitivity was 0.949 and the specificity was 0.876,which was obviously superior than the radiologist's separate diagnosis performance(the sensitivity was 94.5%,the specificity was 66.7%).5.Among the patients with subsolid nodules who underwent surgical resection,the specific pathological subtypes confirmed 823 cases(31.5%)of benign lesions,916 cases(35.0%)of preinvasive lesions and 875 cases(33.5%)of invasive lesions.According to the general clinical characteristics and pathological subtypes,the Macro-F1 value of CNN model established by training in the validation set to predict the benign and malignant degree of pulmonary subsolid nodules is 0.770.Conclusion 1.Doctors' judgment of the benign and malignant lung subsolid nodules based on clinical experience better than simply based on nodule diameter alone,but the difference between different departments and different doctors is large and the overall recognition level still needs to be improved.2.Based on the chest CT,the CT quantitative radiomic model established by MIM graphic workstation can determine the benign and malignant of subsolid nodules and has better recognition performance than some clinicians.3.The established CNN model can effectively identify the benign and malignant of subsolid nodule,and its recognition efficiency is better than that of doctor's diagnosis model and CT quantitative radiomic model.4.The CNN model and the radiologist's judgment are complementary,and the fusion model can better identify benign and malignant of subsolid nodules.5.The established CNN model can predict the benign and malignant degree of pulmonary subsolid nodules,and its effect needs to be further improved. |
PDF Full Text Request