Research And Implementation Of Automatic Pulmonary Nodule Detection Algorithm In Pulmonary CT Images

Lung cancer is one of the cancer with highest morbidity and mortality worldwide.Early de-tection of pulmonary nodules in low-dose computed tomography(CT)scans is very important for the early treatment of lung cancer.In the recent years,the cancer screening with CT scans has been put into practice,but analyzing large numbers of CT scans is a considerable burden for radiologists.So,this thesis proposes an automated and accurate pulmonary nodule detection method,which is based on deep 3D residual convolution neural networks.In this thesis,the automated pulmonary nodule detection method consists two stages.In the first stage,a pulmonary nodule candidate recommendation algorithm is designed to screen the entire CT scan and recommend candidate locations.The aim of the step is to recommend nodule candi-dates at a very high recall and reduce the detection search space.In the second stage,a pulmonary nodule false positive reduction algorithm is designed to reduce the false positives and improve the detection performance.In the pulmonary nodule candidate recommendation algorithm,this thesis designs a deep 3D residual full convolution neural network(FCN)to segment pulmonary nodules and proposes a adap-tive cross-entropy loss to balance the object and background samples.Based on the nodule seg-mentation image,this thesis further proposes a adaptive nodule locating algorithm to recommend the candidates accurately.This pulmonary nodule candidate recommendation algorithm achieves a recall of 98.7%in the LUNA16 dataset and only recommends 154.6 false positives per scan.In the pulmonary nodule false positive reduction algorithm,this thesis designs a deep 3D resid-ual convolution neural network(CNN)to reduce the false positives.Specifically,in the network,this thesis designs a spatial pooling and cropping layer to extract multi-level contextual information of CT data.Moreover,this thesis employs an online hard example selection strategy in the train-ing process to solve the hard/easy sample imbalance problem.The entire automated pulmonary nodule detection method reaches a high detection performance(e.g.,a recall of 85.4%at 0.25 false positives per scan),which outperforms the previous state-of-the-art methods.