Bone Lesion Detection Using Attention Mechanism And Multi-Scale Feature Fusion Network

Orthopedic disease is a kind of common human diseases with widespread and multiple occurrences.In particular,the population structure of our country tends to be aging,leading to a continuous increase in the prevalence of orthopedics.At present,diagnosis of most orthopedic diseases relies on medical imaging,while traditional methods require doctors to perform manual diagnosis,which brings certain risks to doctors and patients.In recent years,more and more deep convolutional neural networks have been successfully applied to medical imaging research,and have shown clinical value in multiple directions.However,due to the variety and complexity of bone lesions,accurate detection of bone lesions in medical images is still a challenging task.To this end,this work will focus on the segmentation and lesion detection of the spine bone,pelvis and other parts in the CT cross-sectional scan image.The main work is summarized as follows:（1）Research on automatic segmentation of bones in CT imagesIn order to extract bone tissue from CT images for analysis and diagnosis,this chapter discusses the bone segmentation technology based on digital image processing.According to the medical image features of bone tissue and the intensity distribution range of its CT value,a segmentation method combining iterative threshold segmentation,region growing and morphological operation is proposed.Firstly,convert the original data into CT intensity,and then normalize,reduce noise,and obtain the region of the human body.Secondly,a method of image enhancement based on gamma transform is proposed.Then,iterative threshold method is used to segment the CT intensity data to obtain these approximate regions of bone,automatically.Finally,select suitable seed points from the approximate area of the bones and segment them by the growth method,and perform post-processing operations.（2）Detection of bone lesions in CT imagesAiming at the problem that it is not easy to automatically detect bone lesions in CT images,we proposed a multi-scale detection network that integrates attention mechanisms module and feature fusion network based on the Cascade R-CNNNetwork.First,according to the characteristics of bone medical imaging and the characteristics of lesion annotation in the data set,more original bone lesion images were screened and a large amount of amplification was carried out.Secondly,through experimental analysis,the most favorable three scale features are selected for multi-scale detection,a fusion method for this work is proposed.Then,the attention mechanism module is used to improve the basic unit of the existing feature extraction network.Finally,dense anchor frame and random multi-scale strategy are used to cover different sizes of target lesions as much as possible.The results show that: on the one hand,compared with the Mimics system,ours bone segmentation method in CT image has better segmentation effect,which has certain clinical significance.On the other hand,the detection network proposed in this work achieves 0.879,0.402,0.815 and 0.417 in bone lesion detection of Recall,AP,AP@0.5and small target index AP@Small respectively.The universal CT lesion detection models3 DGA,ElixirNet,3DCE and the method proposed by Wu Hao（2020）have a recall of0.70,0.574,0.75 and 0.747 for bone lesion detection,respectively.It can be seen that the detection performance of our model has been significantly improved for bone lesions.