Research On Classification Method Of Deep Learning For Minimally Invasive Adenocarcinomas And Invasive Adenocarcinomas From Low-resolution CT Images

Lung cancer has become the first cancer incidence and mortality in the world.As the early symptom of lung cancer,pulmonary nodule has been increasingly found in routine CT screening.Pulmonary minimally invasive adenocarcinomas(MIA)and invasive adenocarcinomas(IAC)are malignant lesions,but their surgical treatment and postoperative treatment vary greatly.At present,there is no study on classifying histological subtypes of lung adenocarcinomas using deep learning.Considering that the low dose,low intensity and popularity of low-resolution CT scanning images with 5 mm slice thickness,and a higher possibility of malignancy from patients with mixed ground glass nodules(mGGNs),we proposed a DenseNet-based deep learning classification scheme for MIA and IAC from low-resolution CT images.The innovations in this paper are the histological subtype classification of lung adenocarcinomas using deep learning for the first time,and the targeted research of the classification on low-resolution CT scanning images with 5 mm slice thicknessBased on the review of intelligent diagnosis methods for pulmonary ground glass opacity nodules(GGN)in recent years,we found the imaging signs were closely related to benign and malignant GGN.Intelligent diagnosis techniques,including deep learning,are helpful in classifying and distinguishing benign and malignant GGN.But there is no research on classify histological subtypes of lung adenocarcinomas using deep learning at present.In order to solve the classification problem of lung adenocarcinoma images with low resolution,we developed the data collection criterion of pulmonary GGN firstly.Based on the criterion,we collected 105 low-resolution CT images with 5 mm slice thickness showing mGGNs from lung adenocarcinoma patients.And these patients were diagnosed as 34 MIA and 71 IAC pathologically.We extracted 2D and 3D nodule samples from the above CT images.Then we built the specific 2D DenseNet and 3D DenseNet deep learing models to explore the diagnostic value of histological subtypes of lung adenocarcinomas on low-resolution CT images using deep learning,especially the DenseNet model.We used five performance indexes including the area under the receiver operating characteristic curve(AUC),accuracy,precision,sensitivity and specificity to evaluate the classification performance.The results indicated that the performance of 2D DenseNet model was not only superior to that of 3D DenseNet model,but also better than that of other four deep learning models.Experimental results demonstrated that deep learning can help to classify MIA and IAC from low-resolution CT images,especially 2D DenseNet model.It can also assist radiologists in predicting histological subtypes of lung adenocarcinomas from low-resolution CT images,and provide guidance for surgical treatment and prognosis judgement on lung cancer patients.