Pulmonary Adenocarcinoma Detection Based On Convolutional Neural Networks

As air pollution gets worse,people are paying more attention to the health of the respiratory system.Lung adenocarcinoma(LA)is a lung cancer that is more common in women or non-smokers.Since the 5-year survival rate is particularly high after surgery of its early stages,early detection of LAs is the most promising path to increase the chance of survival for patients.However,it takes radiologists tens of minutes to diagnose LA in computed tomography(CT)images.Naturally,a fast and accurate Computer-Aid Detection(CADe)system for LA is urgently desired.So far there is little research about CADe systems of LA,but the problem is like lung nodules detection,which has been studied for a long time.The lung nodule CADe systems can be divided into two stages: the stage of detection and the stage of diagnosis.In this research,the former highlight regions of interest(ROIs,areas containing nodules),the latter is to reduce false positives and classify true positives into adenocarcinoma in situ(AIS)and micro-invasive adenocarcinoma(MIA).In CADe systems of lung nodules,convolutional neural networks(CNNs)based CADe systems have better performances than systems based on hand-crafted feature engineering.CNN-based CADe systems diagnose LAs in two stages by using one network for detection and another following network for further classification.Besides costly construction,two-stage systems can only find a single-scale location including many surrounding contexts for each LAs,which leads to a limited accuracy when classifying small LAs.In addition,the LA dataset in this study lacks bounding-box labels which are necessary for localization.This research proposes solutions to the issues above,studying works are as follows:1.In response to the problem that two cascaded CNNs occupy too much memory,the single concatenated path is designed and led to the CNN in the diagnostic stage.The single concatenated path reduces the parameter number by replacing learnable parameters with operations of inversing and concatenating feature maps.This CNN performs better than 3D ResNet in the false positive reduction of lung nodules.2.In order to solve the problem that single-scale location contains many context information and leads to a poor performance in detection of small LAs,feature pyramid structure is involved to the CADe system of LAs.It can generate multi-scale feature map,where small LAs can be tightly encompassed by more precise localization.Based on this,a one-stage CADe system of LAs is realized,and achieves higher sensitivity than two-stage methods in LA detection.3.To get the bounding-box labels of LA dataset,transfer learning is applied to generate the absent size information of LAs.Since the texture of LAs are similar to the lung nodules' texture information,a CNN for LA semantic segmentation is designed and trained on lung nodule dataset.This model can generate the contours of LAs.Finally,the experiment results are analyzed from the distribution of false positives,the drawbacks of bounding-box regression in 3D small object detection and imbalance between different classes.The discussion may provide ideas for following related research.