Comparison of segmentation algorithms for detection of hotspots in bone scintigraphy images and effects on CAD systems

Abstract

The bone scintigraphy scan is one of the most common diagnostic procedures in nuclear medicine for detection of bone metastases. All involvements detected by the imaging process are called hot spots, whether or not they indicate metastasis. Our major concern is the successful segmentation of hot spots, which affects the accuracy of CAD systems developed for the detection of bone metastases. This study examined the extent to which segmentation algorithms can affect the success rate of CAD systems, both in terms of time and of making a correct decision. There is no perfect segmentation algorithm that will provide excellent results for all image types. Using a system developed by the authors, the present study compared several segmentation algorithms for detection of hot spots. Three algorithms known to provide the best results (FCM, SOM, and LSAC) were examined for all details and from different angles. Their performance was measured as 54%, 79%, and 88%, respectively, confirming LSAC as the most appropriate segmentation algorithm. Data obtained by application of the segmentation algorithms were used as input to an artificial neural network model, and the accuracy of the CAD system was measured for each segmentation method. CAD system accuracy rates are 92.3% for LSAC, 86.93% for SOM and 84.62% for FCM. The tolerance of other the segmentation algorithms (FCM and SOM) was measured with reference to LSAC, returning error rates of 7.68% for FCM and 5.37% for SOM. In experimental studies of a total of 706 pelvises, whole body, and chest images, results indicate that more successful segmentation increases the accuracy of CAD systems.