Research On Fracture Detection Algorithm Of Hand X-ray Image Based On Deep Learning

Deep learning is an important branch of artificial intelligence that involves creating computer models to perform various visual tasks,such as object detection and image enhancement.The tremendous success of deep learning in the field of computer vision inspires many researchers to apply it to medical image analysis.A typical example is the use of deep learning-based methods to detect fractures in X-rays.A common and frequent bone disease is hand fracture,for which radiologists typically rely on X-ray examination to evaluate the patient’s condition.As a result,radiologists face a substantial burden of image interpretation due to the large volume of hand X-rays generated.Moreover,due to the limitations of radiologist’s experience and knowledge,missed diagnoses and misinterpretations are not uncommon.Deep learning models have the potential to greatly improve clinical diagnosis of hand fracture,given their efficient and objective characteristics.However,their applications also face several pressing issues that need to be addressed: 1)The lack of a specific medical dataset for hand fractures.The establishment of a specific medical dataset is limited by factors such as imaging equipment and patient privacy,and requires professional medical personnel for annotation,making it more expensive and difficult to obtain compared to natural image datasets.2)The preprocessing methods for X-rays enhancement require further improvement.In preprocessing,manual methods or traditional image processing methods are often used.The former introduces more human factors,while the latter has fixed parameters and lacks flexibility.3)The effect of small fracture detection needs to be improved.The bones in the hand are dense and delicate,resulting in a high proportion of small fractures.Less information is available for these fractures,and the effect of mainstream deep learning methods for their detection need to be improved.The problem of hand X-ray fracture detection based on deep learning is studied in this thesis,which includes research on generative adversarial network(GAN)-based automatic X-ray preprocessing algorithm and attention mechanism-based X-ray fracture detection algorithm.The design of an interactive interface is also included in this thesis,which integrates the research and other basic functions.To solve the problems above,a dataset including 4344 hand X-rays with fractures is established.Based on this dataset,in-depth research is carried out on the automatic X-ray preprocessing and fracture detection models,mainly including the following two contents:(1)Research on GAN-based automatic X-ray preprocessing algorithm.As a preprocessing step for fracture detection,this part uses GAN to translate the original X-rays to the target images,which are manually enhanced under the guidance of doctors.To prevent the loss of hand bone details in the preprocessing,a two-branch GAN that includes a main branch and a detail branch is proposed in this thesis.During the enhancement process of X-ray,the main and detail branches are respectively fed with the X-ray and its high-frequency information,and feature fusion from coarse to fine is performed to maintain the sensitivity of the network to bone detail features.Furthermore,aiming at the problem of difficult optimization of two-branch structure,a mixed loss function that considers both pixel-level and structural similarity is proposed in this study.Validated by the dataset collected in this thesis,better performance in the indicators of mean absolute error,peak signal-to-noise ratio,and structural similarity is achieved by the proposed method compared with the mainstream image translation methods.(2)Research on attention mechanism-based X-ray fracture detection algorithm.Building upon the enhanced X-rays,this research aims to improve the performance of hand fracture detection models.On the one hand,a feature extraction network that combines group convolution and triplet branch attention mechanism is proposed to obtain more accurate fracture features.In the process of feature extraction,grouping the convolution kernels can effectively improve the performance of deep networks and reduce the number of parameters.Additionally,the overlapping imaging of bones and muscles in X-rays causes their features to mix in the same dimension.The triplet branch attention mechanism enables the network to observe the features of X-rays in multiple dimensions,thereby improving the understanding of bone fractures.Attention mechanism also allows the network to better analysis the context information of small fractures,resulting in a more beneficial feature representation for small fracture detection.On the other hand,to avoid missing fracture detection,the post-processing method of filtering out redundant detection boxes is improved.Validated by the collected dataset,the proposed method can achieve higher average precision compared to the stare of the art methods,and has higher application value in the actual clinical environment.