Research On Abnormality Detection And Analysis Methods For Bone Scintigraphy

Bone scintigraphy is a whole-body medical imaging examination.It has been widely used since it is able to detect tumor and bone metastasis earlier than other methods.The diagnostic work of physicians on bone scintigraphy comprises of detection,recognition and analysis of hotspots.In fact,due to the poor image quality and physicians’ experience,different physicians may not reach agreement on the diagnosis of the same patient.This will affect the subsequent treatment of the patient.Computer-aided diagnosis(CAD)system for bone scintigraphy is based on the algorithms of image processing and machine learning.It is able to analyze the bone scan images,detect the hotspots in it,segment out the hotspots accurately and conduct quantitative analysis.The CAD system can not only decrease differences between different physicians and manual workload,but also improve the diagnosis accuracy and sensitivity.At the same time,with the increasing volume of bone scan images,physicians need a image retrieval system to find out corresponding part and similar hotspots from case database and to offer reference for them.In this paper,in view of low signal noise ratio and weak boundary of bone scan images,convolution sparse coding is applied to extract deep features to represent details of bone scan images.The author trained a classifier with AdaBoost,and used the classifier to diagnose if there are hotspots in a certain bone scintigraphy.Similar with physicians,a 3-point interpretation scale is used to score bone scan images.The probability map of hotspots is then constructed by multiple instance learning algorithms.Finally,local signed difference level set is performed to segment hotspots from bone scintigraphy automatically.The experimental results demonstrate that the proposed diagnosis and segmentation methods are effective.Since automatic segmentation is not perfect,the author raised up an innovative interactive segmentation algorithm for hotspot segmentation.Physicians can accurately segment a hotspot with merely offering one point.In this paper,mean field inference is used to propagate information from input points to their neighborhood.At the same time,segmentation is performed with level set methods.The author also proposed an approach utilizing Graph Cut that allows physicians to improve segmentation result.Experiment shows that the proposed interactive approach can segment hotspots from bone scan images easily,fast and accurately.Besides,in this paper,a new content-based bone scan image retrieval system is built for physicians to search similar cases.The author has showed that traditional SIFT feature and Bag of Words model is not capable to represent bone scintigraphy well.In the paper,the author substitutes it with deep features.To keep the effectiveness and accuracy,the author uses kernelbased supervised hashing to construct index for bone scintigraphy database.As shown in the experiments,with the help of the retrieval system,physicians can find similar cases from the database effectively.