Evaluation of ten automatic thresholding methods for segmentation of PET images

Objectives An automatic method for tumor segmentation over PET images has not yet been standardized. Our objective is to test several automatic thresholding methods, standard for other 2D image applications, for segmentation of small spheres acquired on a MicroPET.

Methods Six spheres of different diameters (8-25mm) filled with FDG were scanned separately in a Philips Mosaic MicroPET. Ten automatic thresholding methods were studied, which calculate an optimum threshold for each image based on the exploitation of: histogram shape (Ramesh), space clustering (Otsu, Ridler, Lloyd, Yanni), histogram entropy (Sahoo, Pun) or image attribute information (Hertz, Tsai, Huang). These algorithms were implemented for 3D images and compared to a fixed threshold of 40% of maximum (Matlab). Six spheres with the same diameters were simulated to be used as gold standard. Evaluation was performed quantitatively using 2 metrics, misclassification error (ME) and volume error (VE), taking values between 0 (perfect segmentation) and 1.

Results Segmented images were analyzed visually and the sphere was detected in all cases. Table summarizes the quantitative evaluation of each method, with results arranged in ascending order of average of metrics. According to this average, the six first methods are superior to a standard threshold of 40%. Tsai method showed the best performance.

Conclusions The automatic thresholding methods Tsai, Otsu, Ramesh, Ridler, Hertz and Lloyd (which calculated an optimum threshold for each image) provided superior results than the fixed-threshold of 40% and are ideally suitable for precise tumor segmentation for better diagnosis and therapy planning

• Download figure
• Open in new tab
• Download powerpoint