Deep Learning Based CT Image Detection Of Lung Nodules

As one of the most common and deadly malignant tumors,lung cancer requires timely diagnosis to improve the patient’s cure rate and survival rate.In the medical diagnostic process,relying on radiologists’ manual distinction,the heavy workload and subjectivity of manual diagnosis have led to low detection accuracy,long diagnosis time,and other problems.With the continuous improvement of technology,computer-aided diagnosis technology has become increasingly widespread and in-depth in medical diagnosis.In the lung nodule detection process,CT slices are processed one by one,similar to the doctor’s screening method,but there are some problems that lead to poor detection accuracy: 1.Some nodules in the lung CT are close to other tissues and will interfere with nodule detection;2.Nodule shapes are irregular,and features are not easy to extract,resulting in low detection efficiency;3.The morphology of nodules and other tissues is similar,which can easily cause misjudgment of nodule detection.This paper focuses on the above issues and conducts relevant research on lung nodule detection technology.To address problem 1,this paper designs a pseudo-color preprocessing process based on morphological image processing methods to enhance the characteristics of the nodule area.The experimental results show that this method can increase the m AP value by 13.09%under the same conditions.To address problem 2,this paper proposes a nodule detection network based on deformable convolution,which extracts features of irregular nodules through the adaptive deformation of convolution kernels.At the same time,to address the problem that the size range of lung nodules is large and small nodules account for the majority,this paper optimizes the setting of region proposal boxes and uses a feature pyramid structure to adapt to the needs of lung nodule detection.The best optimization scheme was determined through experiments,and the results showed that the m AP value of the optimized model with deformable convolution can be improved by 1.24%.To address problem 3,this paper introduces attention mechanisms and designs two network structure optimization schemes based on attention mechanisms: 1.combining the attention module with the residual module to optimize the feature extraction process of the overall model and enhance the network’s feature representation ability;2.inserting the attention mechanism module after the deep convolution layer of the feature extraction network to adjust the overall features of the nodules.The experimental results show that different attention mechanisms can improve lung nodule detection in different dimensions,with the highest increase in m AP value being 3.1%.Based on the above optimization schemes for lung nodule detection,this paper proposes a multi-scale lung nodule detection network with fusion channel-spatial attention.The network was trained and verified on the Lung Nodule Analysis 2016(LUNA16)dataset,and the m AP value of the network reached 83.08%.Compared with other mainstream detection networks,it performed outstandingly and achieved good detection results.Figure[39] Table[13] Reference[81]...