Approaches On Neural Network Methods For Skeletal X-ray Image Analysis

Establishing the biological profile and individual identification are critical steps in the fields of forensic medicine,archaeology,and anthropology.Due to the stable morphology,easy observation and corrosion resistance,when the remains are highly decomposed,burned,or skeletonized to the extent that DNA,fingerprint identification,or other methods cannot be used to confirm the identity of the remains,bone and its X-ray images,CT images are often used for analysis or compared with images in a massive database to assist in individual identifying or estimating the biological profile of the remains,such as sex,age,and ethnicity.Traditional methods mainly rely on manual measurement which heavily depends on the accumulated experience,uniform standards,and meticulous training of researchers.With the widespread application of deep neural networks in fields such as object detection,image recognition,feature extraction,and image generation,artificial intelligence-assisted forensic research can help researchers to process large amounts of skeletal images more accurately and quickly and perform automated analysis with few subjective biases and errors.Against this background,this paper proposes to solve specific application tasks of skeletal X-ray images in forensic medicine by neural network,including the detection and sex estimation tasks of skeletal X-ray images,as well as feature extraction and skeletal generation tasks.The main research content and contributions are summarized as follows:This paper takes the femur as an example to study the tasks of object detection and sex estimation of skeletal X-ray images,and proposes an automated framework for femur proximal detection and sex estimation.Firstly,the collection of femur proximal X-ray image dataset is transformed into the extensive collection of pelvic X-ray images and the detection of proximal femur.Secondly,a targeted data augmentation strategy is designed for sex estimaton tasks of femur proximal X-ray images,which greatly enhances the robustness and generalization of the network.Finally,a classification network is trained based on transfer learning theory to complete sex estimation of femur proximal X-ray images.The results show that the proposed method achieves competitive performance,with an accuracy of 94.6%,which is close to the best accuracy of current manual methods.The class activation mapping method is used to visualize the sex estimation results,and it is observed that the model has a correct understanding and decision-making basis for the task.This paper presents a study on the feature extraction and bone generation tasks of bone X-ray images using the chest skeleton as an example.We propose a multi-view chest X-ray adaptive bone feature extraction and generation framework based on conditional generative adversarial networks,which can generate bones from a single standard chest X-ray image.This framework is the first to consider multi-view bone feature extraction and generation without any manual annotations for the chest.Firstly,the proposed method fully utilizes the multiview chest X-ray images obtained by dual-energy subtraction,and constructs two cross-view datasets based on the rules designed according to the contrastive learning theory.Secondly,a Siamese/triplet feature extraction network is designed with octave convolutions to learn the shared bone information in the cross-view dataset.Finally,a bone generation network is designed based on the conditional generative adversarial network to generate bone images using the learned shared bone information.The results show that the proposed method can generate the bone image ignoring the overlap of multiple anatomical structures in the chest radiograph,with RMAE of 3.45% and a FID of 1.12,which preserves bone details while reducing the distribution shift of image features,and obtains the highest manual score.In summary,the proposed method can help researchers to automatically analyze bone Xray images.It is believed that with the continuous development of neural network,artificial intelligence can be more widely applied in forensic medicine,anthropology and archaeology or other related fields,contributing to the traceability of human history,the development of human society,and the progress of human civilization.