Research On Wrist Bone Age Assessment Based On Deep Learning

Bone age can be used to assess the growth and development stage of bones by their shape,size and other characteristics,which is called bone age assessment.Through the gap between bone age and real age,it can be known whether bones are developing normally.Bone age assessment has been widely used in clinical medicine,forensic medicine and athlete selection.Left hand X-ray image is the most commonly used data for bone age assessment.And a complete assessment system has been set up,among which the internationally used methods are GP atlas and TW scoring.However,these methods are complicated and time-consuming.Therefore,the technologies and methods of computer-aided bone age assessment have been paid more and more attention by researchers.With the rapid development of deep learning and neural network,bone age assessment based on deep learning has gradually become the focus of most researches.The main reason is that data processing is simple,model is robust and the result is of high accuracy.Some studies have pointed out that wrist is the most significant area for age assessment in the low age range and it can reach better accuracy compared to the whole hand image in this age range.Moreover,wrist bone age plays an important role in the diagnosis of growth hormone deficiency and other diseases.In addition,age assessment based on wrist can reduce the exposure area of hand in X-ray.Local features play an important role in bone age assessment task.Some scholars mainly obtain local features by artificial selection,detection network or heat map.Therefore,this thesis focuses on designing efficient global and local feature extraction network.The main contents are listed as follows:(1)Due to the effect of background,some existing models fail to segment small bones.To solve the problem,this thesis designs a carpal bone segmentation network U2 Net,which uses multi loss function to suppress the influence of the background,and achieves fine segmentation by dual decoder.Compared with UNet and UNet ++,it achieves better result in the test dataset.(2)The main basis for bone assessment by expert is the outline,shape and size of bones.Therefore,a bone age assessment network Boception with strong global feature extracting ability is proposed.By taking the carpal mask as the auxiliary input,the network pays more attention to the carpal region.Its feature fusion and residual spatial attention module can make great use of the high-level abstract semantic information and the low-level shape and location information,and focus on the key region.Finally,by using cosine decay learning rate strategy and L1 loss function,the model reaches a mean absolute error of 6.4 months in wrist dataset.(3)In order to increase the diversity of global features for better results,most researches on bone age assessment are devoted to finding important local regions of input image to get supplementary information,but these methods have some problems such as time-consuming and complex process.In this thesis,a Muti-part Boception Network(MBNet)is proposed to combine global features and local fine-grained features.The MBNet creatively introduces a feature segmentation branch and an attention-based channel splitting branch.These two branches can extract potential and obvious local features from global features respectively.Experimental results show that the mean absolute error of the proposed model in the wrist dataset is 5.9 months,which is 11% lower than that of the Inception-v3,and the mean absolute error of the model in the public hand dataset is 5.0 months.(4)Based on the work of this thesis,a practical exploration is carried out.Actually,a bone age assessment software based on Py Qt is developed.The software can automatically mark the wrist area,segment the carpal bone,and assess the bone age by hand image and wrist image.