Research On Skeletal Abnormally Detection Method Based On Deep Neural Network

Bone disease is one of the five most common diseases,plagues nearly 1.7 billion people worldwide,and is the most common cause of chronic pain and physical disability.In recent years,researchers have conducted in-depth research on the detection and segmentation methods of medical images,and have achieved fruitful results,but there are few studies on abnormal diagnosis and detection methods for bone lesions.In this context,the main research contents of this article are as follows:1.This paper proposes a convolutional neural network structure that combines shallow texture features to diagnose skeletal images.In radioactive bone images,there are problems such as unclear dividing lines between tissue structures,little difference in texture between the lesion area and the normal area.For this,this article first uses texture analysis to count the local texture features in the region of interest,and then the texture features the image and the original bone image are fused,and then the deep neural network is used for classification.The experimental results on the large-scale radioactive bone abnormal classification data set MURA show that the model combined with texture features can better diagnose abnormalities of bone images.2.This paper proposes a two-stage method for diagnosis of abnormal bone images.Since bone diseases are mainly composed of abnormal objects,degenerative arthritis,fractures and injuries,etc.,the lesions that cause skeletal imaging have the problems of random location,unstable size and shape,and large differences in types.In this regard,we learn from Boosting’s learning ideas,first use a classification model to quickly identify easily distinguishable lesions,and then in the second stage for fine recognition of difficult to distinguish lesions.The experimental results show that the twostage method proposed in this paper can better diagnose the bone image with less obvious lesion area.3.For the first time,this article completes the precise positioning of the lesion for the MURA dataset.The computer-aided diagnosis technology not only needs to judge whether the bone image is abnormal,but also needs to accurately locate the focus area.At present,there are few publicly available skeletal lesion detection data sets,so we first calibrated the lesion position on the MURA dataset,and then based on the Faster RCNN,SSD and YOLO-V3 model framework,the lesion position was achieved by adjusting the model structure For the detection function,on the MURA dataset,we achieved a map value of 77.97%.