Research On Rotating Target Detection Algorithm For Medical Diagnosis

With the intensification of population aging and normalization of epidemic management,people’s health awareness is becoming stronger and stronger,and regular physical examinations will be organized.Medical imaging examination such as CT(Computed Tomography)and MRI(Magnetic Resonance Imaging)are important items in physical examination because of its high resolution of the lesion.Nowadays,lesion detection mainly uses target detection technology.However,medical images are mainly gray images,which have the characteristics of blurred target boundaries,large differences in lesion proportion and large background interference,resulting in the problems of small target missing detection and multi-target missing and wrong detection in the existing medical image target detection methods.In this regard,the rotating target detection algorithm for medical diagnosis is proposed,aiming to improve the accuracy of lesion detection.The specific work content is as follows.(1)Aiming at the problems of missing detection of small targets,single direction detection and aliasing of targets in single lesion detection,a single lesion rotation detection algorithm is constructed based on non-local attention mechanism.In view of the quick connection and identity mapping of residual networks,the backbone network is designed to obtain the global dependency of any two pixels in the feature map through nonlocal attention mechanism,improving the detection effect of the network for small targets.An Angle sensitive module is designed to identify rotating lesions in multiple directions by introducing rotation angles to make up for the deficiency of horizontal bounding boxes in multi-angle target detection.For single-type and multi-type target aliasing problems,RNMS(Rotated Non-Maximum Suppression)and MRNMS(Multi-type Rotated Non-Maximum Suppression)modules are constructed.By sorting,confidence comparison and calculating Io U(Intersection over Union),the bounding boxes with Io U higher than 0.7 is filtered to eliminate aliasing and achieve accurate detection of lesion location.(2)In view of the problems of multi-target wrong detection and missing detection,the deviation of key feature points of the bounding boxes and the large loss of classification and regression in the detection of multiple lesions,a multiple lesion rotation detection algorithm is proposed based on multi-head attention mechanism.Eight-head attention mechanism is designed to learn feature information in different representation subspaces to improve the accuracy of multi-target detection.The bilinear interpolation algorithm is used to fit the pixel value and position information of the key feature points to reduce the deviation of the key feature points of the horizontal bounding boxes and the rotating bounding boxes.A multi-class loss function is designed,which combines the focal loss function and Smooth L1 function to increase the weight of complex and misclassified samples.Additionally,this approach reduces the loss of easier-toclassify samples,hence improving the accuracy of bounding boxes regression.These modifications result in a more robust algorithm that is better equipped to handle the detection of multiple lesions with higher accuracy rates.Through the experimental results of Rotated PET-CT-Dx Dataset and Rotated Lesion Dataset,the m AP of the detection scheme for single lesion in this thesis reaches 87.3%,and the m AP of the detection scheme for multiple lesions reaches 85.7%.The experimental results show that the single lesion rotating detection algorithm based on non-local attention mechanism can obtain the correlation between pixels and improve the response of the network to small targets.The multiple lesion rotating detection algorithm fused with multi-head attention mechanism extracts the potential features of the target from multiple dimensions,and effectively reduces the multi-target wrong detection and missing detection phenomenon.