Automatic Recognition Algorithm Of Radionuclide Bone Scan Image

Whole-body bone scan image analysis is often used to assist nuclear medicine physicians in detecting cancer bone metastases.The intensification of aging and the rising population have led to an increase in the demand for diagnostic testing of whole-body bone scans.Nuclear medicine physicians must carefully read bone scans and write diagnostic reports,which have led to a significant increase in their workload.Based on the characteristics of bone scan images,this thesis adopts a new method of machine learning to study the automatic identification technology of radionuclide bone scan images to assist clinicians in diagnosing bone metastasis.It is required to automatically mark suspected bone metastases and abnormal parts.The main contents and research results of this article are as follows:1)Segmentation of whole-body bone scan images based on anatomy.Combined with the biological characteristics of the bone scan image,the bone scan image is divided into 26 homogeneous sub-regions,including the head,shoulders,chest cavity,elbow joint,etc.The whole segmentation process includes the calculation of feature values based on statistics and the positioning of feature points.2)A method of relocating image segmentation is proposed.According to the correlation between feature values,feature points,and empirical parameters,a small number of erroneously segmented bone scan image sub-regions are repositioned,which significantly improves the image segmentation accuracy.3)A spinal contour segmentation algorithm is proposed.After extracting the outline of the spine by the optimal threshold method,the outline of the spine generally includes organs such as kidneys and ribs with high-gray.Using the spinal contour cropping algorithm proposed in this paper,the non-vertebral contour parts can be trimmed to accurately extract the spinal region.4)Construct a deep parallel learning network to detect bone metastases and abnormal regions.The 26 homogeneous sub-regions generated by the segmentation are simultaneously input to 26 parallel deep learning networks to detect anomalies in each sub-region in parallel.Parallel deep learning networks can quickly and accurately detect and locate abnormal regions in whole-body bone scan images,and the detection speed is much faster than experienced nuclear medicine experts.This subject experiment segmented 689 whole-body bone scan images with a success rate of 97.2%.For the parallel deep learning network,sensitivity,specificity,and AUC values were adopted as performance indices to evaluate the proposed model.The sensitivity and specificity of a single deep learning network are up to 97.3% and 99.9%,and the highest AUC value is 99.0%;the sensitivity and specificity of the overall parallel deep network are 71.8% and 99.2%,and the AUC value is 86.2%.The entire network can detect a patient’s bone scan image in milliseconds.Experiment shows that the parallel deep learning network proposed in this paper has high specificity and acceptable sensitivity,and is fully valuable for clinical diagnosis.