Researches On Anomaly Detection And Recognition Algorithm Of Bone Scan Images Based On Deep Learning

As a nuclear medicine imaging examination,bone scan has the advantages of early detection of lesions and low cost,and is one of the most common methods to detect tumor bone metastasis.The diagnosis of bone scan images relies heavily on the medical knowledge and diagnostic experience of physicians,but the image quality is poor and the signal-to-noise ratio is low.With the improvement of people’s health awareness,the difficulty and workload of nuclear medicine physicians have greatly increased.At present,there are some researches of the diagnosis for bone scan images by the assist of computer,but most of them are only based on traditional methods and specific private small datasets,and only a few researches develop bone-scan study based on deep learning technology.Most of them only utilize methods of image classification and assist doctors in analyzing and diagnosing bone scan images according to the classification results.So the methods of bone scan image analysis based on deep learning are worthy of further study.There is no large public data set in the field of bone scan images,and most of the existing deep-learning-based researches only utilize the methods of image classification.To solve this problem,some research had been carried out and the main content of this thesis is as follows:1)Construct the first large publicly available bone scan image dataset.The dataset contains whole-body bone scan images of 2950 patients,and each image is divided into 13sub-regions by region to form a sub-data set for the study of images of specific parts.The dataset is established in the format of the PASCAL VOC dataset for object detection.Each whole-body bone scan image corresponds to a label file.the type and location information of hot spots in the image are saved in the label file.The hot spots are classified into normal and abnormal.Based on this dataset,this paper conducts a series of image classification and object detection verification experiments.The experimental results demonstrate the practicability and robustness of the dataset.2)Propose a hot spot detection and recognition algorithm in bone scan images based on multi-view feature fusion.The algorithm considers the symmetrical relationship between the anterior and posterior views of the bone scan image,and performs image pixellevel fusion or high-level semantic feature fusion in the image input stage,feature extraction stage and decision stage of object detection,and detects hot spots in the bone scan based on the fused image/feature map/feature vector.The experimental results show that the feature fusion of the anterior and posterior views in the feature extraction stage makes the greatest overall improvement for the hot spot detection.3)Propose a hot spot detection and recognition algorithm in bone scan images based on the customized RPN reling on GMM.The algorithm considers the prior knowledge that the hot spots are mainly distributed in the spine,pelvis and other regions on the bone scan image,and utilizes the Gaussian mixture model（GMM）to model the distribution of the hot spots in the bone scan image,and then customizes Region Proposal Network（RPN）according to the Gaussian mixture model（GMM）to detect hot spots.The experimental results show that the customized RPN based on GMM improves the effect of hot spot detection.