Research Of Medical Image Lesion Detection Based On Multi-local Point Feature Information Fusion

With the rapid development of computer vision,object detection research and applications have attracted much attention.In the context of such a high incidence of cancer,automatic detection of lesions based on computed tomography can effectively improve the efficiency of disease diagnosis.Computed tomography images are used as one of the most useful tools for localizing lesions and diagnosis,which can largely assist clinicians in identifying and analyzing lesions based on information such as their location and size for early cancer detection and treatment.Automatic lesion detection plays an extremely important role in the computer-aided diagnosis process,and its lesion detection accuracy can directly affect the risk assessment and treatment of subsequent diseases.Due to the role and advantages of deep learning with data feature extraction and sample law learning,more and more researchers have started to study automatic lesion detection based on deep learning.To address the limitations of the current lesion detection methods,which cannot make full use of local information for bounding box regression prediction and fail to integrate regional features for object classification,this thesis proposes a generalized and high-quality lesion region detection model.The main research contents of this thesis are as follows:(1)Most current detection methods still follow the traditional regression strategy,which predicts the offset of the bounding box centroid to fit the object bounding box or uses multiple extreme points to determine the object location,which is not a good fit for the detection results of lesions and the real lesion area,which is not conducive to the subsequent lesion segmentation and radiation-related lesion morphology estimation operations.For the problems of low efficiency and localization accuracy of bounding box regression,multi-key point regression strategy and enhanced feature pyramid are proposed.The multi-key point regression utilizes multiple local features for global bounding box offset prediction,and it corresponding to the first Io U threshold is combined with a heat score map to improve the regression prediction capability in the case of limited local points.A bottom-up information propagation path is also created in the FPN to effectively improve the transmission and extraction of the underlying geometric location information.The experimental results show that the proposed method can achieve the expected results with higher detection accuracy(AP and AP@75)and sensitivity compared with the current mainstream methods.(2)Most generic lesion detection models do not highlight the information of each category of features in the classification phase,as well as fail to recover and utilize the past experience of processing each category,resulting in the discarding of useful information from other levels in the category prediction process.To make full use of multi-level information,an adaptive multi-point feature fusion module is designed to aggregate regional feature maps from multiple levels,perform operations such as regional attention on candidate regions to highlight object features,and incorporate a feature fusion module in the final object classification process.By analyzing the images of regions containing different kinds of lesions,it is found that the proposed method has a significant improvement in detection results and a good generalization ability to different kinds of lesions,and indicator AP@50 can reach63.6%,to obtain a higher accuracy.