A novel segmentation method for breast ultrasound images based on neutrosophic l-means clustering.

PURPOSE

Fully automatic and accurate breast lesion segmentation is an essential and challenging task. In this paper, the authors develop a novel, effective, and fully automatic method for breast ultrasound (BUS) image segmentation.

METHODS

The segmentation method utilizes a novel phase feature to improve the image quality, and a novel neutrosophic clustering approach to detect the accurate lesion boundary. First, a region of interest is generated to cut off complex background. After speckle reduction, an enhancement algorithm based on phase in max-energy orientation (PMO) is developed to further improve the image quality. The PMO is a newly proposed 2D phase feature obtained by filtering the image in the frequency domain and calculating the phase accumulation in the orientation with maximum energy. Finally, the authors propose a novel clustering approach called neutrosophic l-means (NLM) to detect the lesion boundary. NLM is a generalized clustering method that can be used to solve other clustering problems as well. In this paper, NLM is used to segment images with vague boundaries, and to deal with uncertainty better. To evaluate the performance of the proposed method, the authors compare it with the traditional fuzzy c-means clustering, active contour, level set, and watershed-based segmentation methods, using a common database. Radiologist's manual delineations are used as the golden standards. Five assessment metrics are utilized to evaluate the performance from different aspects. Both accuracy and efficiency are analyzed. Sensitivity analysis is also conducted to test the robustness of the proposed method.

RESULTS

Compared with the other methods, the proposed method generates the most similar boundaries to the radiologist's manual delineations (TP rate is 92.4%, FP rate is 7.2%, and similarity rate is 86.3%; Hausdorff distance is 22.5 pixels and mean absolute distance is 4.8 pixels), with efficient processing speed (averagely 9.8 s per image). Sensitivity analysis shows the robustness of the proposed method as well.

CONCLUSIONS

The proposed method is a fully automatic segmentation method for BUS images that can generate accurate lesion boundaries even for complicated cases. The fast processing speed, robustness, and accuracy of the proposed method suggest its potential applications in clinics.