Research On Automated Lung Nodule Detection Based On Deep Convolutional Neural Network

The automatic detection of lung nodules based on CT images is of great value for the early detection and treatment of lung cancer,and thereby saving the lives of the patients.Early lung nodule detection methods was based on either a rule system or training machine learning algorithm with hand-crafted features.Although many achievements have been made,for a task like automatic detection of lung nodules whose targets varies greatly in appearance,shape,size,and contextual characteristics,not to mention the lots of areas similar to nodules in the background,both of these methods are not only limited in performance but also insufficient in generality.In recent years,with the increase of data volume and the improvement of hardware computing capabilities,the development of deep learning has brought breakthrough progress in many research fields.One of its advantages is that it can automatically learn the features that is optimal for a specific task from a large number of data,hence avoids the tedious work of desigining features manually,and makes the learned features more discriminative and generic.Therefore,more and more researchers are beginning to devote themselves to the combination of deep learning and automatic detection of lung nodules.Although these studies have preliminarily verified the effectiveness of deep learning,in order to better apply deep learning to the automatic detection of lung nodules,many problems and challenges still need to be faced,such as the lack of annotated data,the class imbalance problem and the computational resource consumption,etc.To solve these problems,we starts with designing a simple and effective automatic lung nodule detection method,and proposes corresponding countermeasures from four aspects:data preparation,model structure design,optimization mechanism improvement,and loss function design.These strategies have achieved certain effects,and the performance metric of our overall model has reached 0.899 on the internationally public pulmonary nodule detection data set Lung Nodule Analysis 2016(LUNA 16),and the average processing time on a single CT image is approximately 12 seconds.In addition,compared with other methods evaluated on this data set,our method has a comparative advantage in terms of comprehensive considerations of performance,efforts of implementation,and detection efficiency.For clinical practice,the radiologists are more interested in the detection sensitivity of lung nodules when the average number of false positives per CT image is 1,2,and 4,and our method can achieve 0.916,0.948,and 0.972 respectively,has potential to be applied in clinical practice.