Research On Improved Faster R-CNN Method For The Detection Of Lung Lesion Areas

Lung is the organ with the largest number of pathological changes and diseases in thoracic surgery.Common lung diseases include pulmonary nodules,emphysema and pulmonary fibrosis,etc.,which can lead to respiratory failure with high fatality rate and difficult prevention and treatment.Currently,the early screening of lung diseases mainly relies on two kinds of medical images,X-rays and Computed Tomography(CT).They have high complexity,large amount of data,high requirements for doctors’ experience in pulmonary disease diagnosis,and heavy workload of film reading,which can easily lead to misjudgment and missed judgment.Therefore,it is necessary to realize the automatic detection of lesion areas in X-ray and CT images.However,the complex lung structure and the uneven lung image quality pose challenges for the automatic detection of focal areas in X-ray and CT images.This thesis gives full consideration to the characteristics of X-ray and CT images as well as the characteristics of lung lesion areas,and carries out research on automatic detection methods of lung lesion areas,mainly including the following contents:(1)The existing X-ray detection algorithm does not make full use of the edge and shape information of the lesion area,and the receptive field is single,resulting in missing small-scale lesion area.In order to solve this problem,a multi-scale feature fusion algorithm is proposed to detect focal areas in X-ray images.In this algorithm,a deformable convolution module and a multi-scale feature fusion module are introduced in the Faster RCNN detection network.The deformable convolution module can predict the lesion shape according to the lesion boundary,and the multi-scale feature fusion module has three scale receptive fields.Fusion of different scale features can detect different scale lesion areas.The proposed method was tested on Chest X-box dataset,and the results showed that the accuracy of the proposed method was higher than that of the comparison method.(2)The amount of lung CT image data is large,and the proportion of lesion area on CT is small.The current CT lesion area detection methods have a large number of anchor frames,and the detection speed is slow.To solve this problem,an automatic detection algorithm based on location guidance and shape prediction is proposed.By introducing the position constraint of candidate frames,the algorithm can effectively reduce the number of candidate frames and reduce the amount of computation.By predicting the shape of the anchor frame based on the location information,the anchor frame closer to the size of the lesion area can be generated,further improving the detection accuracy.The proposed method was tested on LUNA16 data set and compared with current mainstream methods.The results show that the proposed method can effectively reduce the detection range of pulmonary nodules and the number of anchor frames generated.The work of this thesis enriches the theory and method of automatic detection of lung disease regions from different dimensions,improves the detection accuracy of different scale disease regions,and has important theoretical significance and practical value.There are 30 images,7 tables,and 99 references in this thesis.