X-ray Bone Age Prediction Method Based On Convolutional Neural Network Transfer Learning

Pediatric bone age prediction plays a key role in many medical and forensic evaluations.Traditionally,people use the Greulich and Pyle method(GP method)or Tanner and Whitehouse(TW method)scoring method to determine the bone age represented by the corresponding hand bones on an X-ray.The GP method relies on manual evaluation of the similarity between the X-ray film and the image in the standard atlas,and uses the most similar standard picture as the predicted bone age.The TW method divides the X-ray film into several key regions,and proposes more targeted evaluation methods and indicators for different key regions.However,the traditional method cannot avoid the difference in personal experience and ability,and at the same time,the evaluation efficiency cannot be guaranteed.Deep learning(DL)usually requires large-scale annotated image data sets,but due to the high cost of medical data acquisition and privacy considerations,it is difficult to obtain a large amount of bone age data.Use Convolutional Neural Network(CNN)to learn for bone age classification and transfer low-level semantic features of hand bone images learned from the Radiological Society of North America(RSNA)data set to Tongji)A small data set collected by the hospital.The performance of the model is measured according to its Mean Absolute Distance(MAD)on the test set.In order to verify the effectiveness of the above model,a public dataset from RSNA and a small dataset from Tongji were obtained.The above data set is input into the model for training and testing its MAD.The experimental results show that the MAD of the proposed network framework in the RSNA dataset is 0.19 years,which is better than the 0.36 years MAD of the first proposed method in the bone age recognition competition held by RSNA.When the network framework transfers learning to a small data set collected by a local hospital,the parameters of the first four layers can achieve the best MAD of 0.169.At the same time,when using transfer learning for gender classification,the accuracy rate of the RSNA dataset is used.It is 93.64%,and the accuracy of using Tongji dataset is 95.67%,indicating the effectiveness of the extracted features.The results of feature visualization show that the bone age features focus on the joint gaps of the hands and the distal radius of the ulna,and pay more attention to the joint gaps of the index and ring fingers than the TW score.Finally,ablation experiments have also proven that gender information is critical for hand bone age prediction.