Design And Implementation Of Multi-site Lesion Detection System Based On Improved Faster R-CNN

High-definition medical CT(Computed Tomography)plays a key role in clinical medical diagnosis.The doctor can make a diagnosis of the patient’s condition and give a treatment plan through the CT image.In the current computer aided diagnosis system,the existing detection function only performs lesion detection for a specific part.However,it can be found that many lesions are actually related in the actual clinical diagnosis.The detection of a single site is not conducive to doctors to make a comprehensive diagnosis of the disease.Multi-site lesion detection can detect lesion metastasis earlier and explore the relationship between different lesions at the same time.In this thesis,a multi-site lesion detection system based on improved Faster R-CNN is established for the limitations of single lesion detection in medical diagnosis.This system can detect lesions on CT images of multiple parts and achieve good detection results.In addition,doctors and administrators can manage patient visit information and doctor personal information in this system,which is convenient for doctors to use this system in clinical diagnosis.The specific work of this thesis is as follows:(1)Pretreatment the CT images in the dataset,including image denoising and image enhancement.According to the analysis and comparison of different denoising methods and enhancement methods,it is determined to use Wiener filtering and wavelet transform to pretreatment the image.(2)The improved Faster R-CNN network is used to train the pretreatment CT images.First,the VGG16 feature extraction network is improved to enhance the resolution of the feature map and increase the sampling rate of the lesion area.Second,the anchor frame in the area suggestion network is redesigned according to the size of different lesion frames in the dataset.The feature maps pass through the area suggestion network to obtain candidate suggestion areas.Finally,the final prediction results are obtained through the pooling layer and the fully connected layer.In view of the fact that the traditional non-maximum suppression algorithm misses the detection of adjacent lesions in the experiment,this thesis introduces a Gaussian weighted penalty function to improve it,and it achieves good results.(3)Use the FROC curve to evaluate the improved detection model.The experimental results show that when the average number of false positive regions per image is fixed,the improved model detection sensitivity is superior to the original model and can detect more real lesion regions.The improved detection model is used to detect lesions in CT images of multiple parts.The experimental results show that this network model can detect lesions in CT images of multiple parts better.(4)A multi-site lesion detection system is established based on the improved Faster R-CNN multi-site lesion detection model using the Django framework in Python.Front-end system uses Bootstrap framework.Use jQuery plugin for page layout.The system contains three functional modules,including login module,detection module and management module.This system can assist doctors to diagnose multi-site lesions and manage patient information in clinic.