Pulmonary Nodule Detection Based On Feature Selection

According to the cancer statistics,lung cancer is the leading death reason in all ages,which seriously threatens our lives.The survival rate of lung cancer depends on when the disease is diagnosed.Researchers show that early diagnosis helps improve the survival rate over 60% with less treatment cost.However,only few diagnosed lung cancers are detected at an early stage for lack of obvious symptoms.Pulmonary nodule detection,therefore,plays an important role in lung cancer screening.Numerous computer-aid diagnose systems,mainly categorized into feature descriptor exploitation,threshold and morphological operation,learning based classification,have been developed for nodule detection.All these algorithms are not practical for two main reasons.Firstly,experiments on tiny dataset cannot credibly verify the generalization ability of the model.Secondly,the morphological variation of nodules is so great that the existing algorithms are usually unable to express the characteristics of all types of nodules.In this paper,we collect a new large dataset,and gradually improve the detection framework from the perspective of both data augmentation and feature selection to tackle the problems mentioned above.We utilize the excellent characteristics of deep learning model and focus on the three key functional modules: data augmentation,preliminary nodule detection and false positives reduction.We make innovative work as follows: 1)crossmodal data augmentation.As we know,fully leaning of model has a great appetite for training data,which cannot be satisfied by a tiny dataset.Additionally,straightforward mixing of datasets captured from different devices has a negative effect on the detection accuracy due to modal differences.We transfer nodule characteristics from source modal to target modal via a cycle-consistent generative adversarial network(CycleGAN).Experimental results demonstrate that our transfer model can improve the accuracy by 2%;2)It is the kernel module of the system to propose nodule candidates from Computed Tomography(CT)images.Firstly,we use the selected hand-craft features to improve the classic Aggregate Channel Feature(ACF)detection framework.It applies to detect nodules in CT images whose slices thickness is 5mm and effectively distinguishes the nodule with other tissues.Finally,the detector achieves 91% accuracy.3)We fully utilize convolutional neural network(CNN)features,making up for the defect of hand-craft features,for hard samples detection.Considering that CT slices with thickness less than 2mm contain more small nodules,we propose to extract meticulous features for small nodules representation by adding a deconvolution layer in Faster Region-based Convolutional Neural Network(Faster-RCNN)model.This model further reduces the miss rate compared to the ACF detector.4)Numerous false positives are maintained after preliminary detection.To further optimize the performance,we construct a strong classifier using a 3D-CNN network,with space features embedded.The module is proved to work well and limit false positives per scan.