Research About Lung Nodule CT Images Classification Based On Deep CNNs

Nowadays lung cancer is one of the most vital malignancies to human health,and the best solution to it is still early diagnosis and targeted treatment.Among various approaches for lung cancer early screening,low-dose spiral computed tomography is widely considered to be the most effective one.However,image representation of lung nodule is very complicated,so radiologists are prone to visual fatigue after working for a long time,which makes it hard to avoid missed diagnosis and misdiagnosis.With rapid development of computer technology especially computer vision,Computer Aided Diagnosis gets widely used in medical image analysis.Through increasing accuracy and stability of diagnosis and reducing time it needs,Computer Aided Diagnosis can significantly improve quality and efficiency of medical image analysis,reduce the chance of wrong diagnosis caused by subjective factors and missed diagnosis brought by human eye oversight.But Computer Aided Diagnosis needs us design image features manually,which not only rely on human’s judgment but also brings a lot of work.Pulmonary nodules classification algorithms in traditional computer aided diagnosis need extract features of nodule image manually,and this makes features extracted lack diversity,which brings the problem that model can’t fully describe features of pulmonary nodules.Recent years deep convolutional neural networks(CNNs)have developed rapidly and excellent performance of CNNs in image feature capture makes it be mainstream algorithm in image classification tasks.What’s more,CNNs don’t need manually designed features.There is a certain lag between CNN architectures existing work uses and CNNs theory in current stage,and accuracies of existing algorithms designed for lung nodule classification are still need to be improved to achieve the practical standard.So on the basis of full investigation of CNNs theory,this paper proposes applying Inception-ResNet and CondenseNet into automatic lung nodule CT image suspiciousness classification to help radiologists screen malignant lung nodules from large-scale lung CT images.Further we introduce self-attention mechanism in SENet into our CNN architecture to picture reliance among feature channels and adjust channel’s relative feature strength adaptively through global loss function.What’s more,considering the 3D nature of lung CT images,we apply 3D convolution kernels into our network to capture information hidden in reliance among sequent slices,which realizes better classification to lung nodule CT images.Introduce self-attention mechanism of SENet and 3D convolution into Inception-ResNet and CondenseNet,and we obtain target CNNs architectures:3D-IR-SENet and 3D-CD-SENet.Further,we combine them together and obtain a better network architecture:3D-HybridNet.Experiment of this paper is based on subset of LIDC/IDRI lung CT image dataset,LUNA16.In this experiment we extracted 48×48 2D lung nodule images and 48×48×93D lung nodule cubes.Results of experiment show that 3D-IR-SENet and 3D-CD-SENet achieve accuracy of 91.27%and 91.64%respectively for classification of lung nodule suspiciousness in LUNA 16 dataset.Compared with initial Inception-ResNet and CondenseNet,lung nodule malignancy classification accuracies of 3D-IR-SENet and 3D-CD-SENet increase 2.79%and 2.58%,respectively;Compared with other algorithms,3D-CD-SENet achieve better accuracy for lung nodule classification task,increasing 0.38%in comparison with the best one,and accuracy of 3D-IR-SENet is slightly better than the highest baseline.What’s more,our hybrid CNN architecture 3D-HybridNet’s accuracy achieves 93.18%,which is 1.92%higher than highest baseline.Experiment results prove effectiveness of convolution neural network architectures this paper chooses and modifications conducted to them.