Medical Bone Fracture Detection Task Based On Deep Learning

Computer aided diagnosis(CAD)provides medical diagnostic proposals to radiologists by computer-suggested approaches,which effectively reduced the workload and improved the work efficiency of clinicians.In the bone fracture detection task,accurate recognition and detection of bone fracture location is the key to computer aided diagnosis.In this thesis,a two-stage object detection network with multiple reception field is proposed,which can automatic and accurate detect bone fractures from X-rays.The network solved the problem of low detection accuracy which caused by scale variance of bone fractures,improved the detection performance.First,the proposed backbone network has three pathways,each pathway generates feature maps with different reception field by adopting dilated convolution with different dilation rates.Feature maps with multiple reception fields are conductive to extract more learnable features,enhance the resolution and build multiple-scale context.Backward connection structure connected feature maps in different depth,which enriched the complexity of features and fully trained the parameters.Second,attention mechanism is adopted based on the proposed multiple reception field network.By adopting attention blocks in backbones,the network reweighted the parameters of the channels in the feature maps and built channel-wised connection.A Refined Feature Pyramid Network structure is proposed based on Feature Pyramid Network(FPN).In Refined FPN,up-sampling pathway is replaced by deconvolution and an attention feature fusion block is designed to replace lateral connection.Feature maps generated by refined FPN can avoid losing too much semantic information while improving resolution.At last,due to small medical dataset,transfer learning is adopted in the proposed network to finetune the parameters from the pretrained model to accelerate and optimize the learning efficiency.By using data augmentation,data samples are expanded on the basis of original data to provide more learnable features for the network.