Ultrasound liver tumour active contour segmentation with initialization using adaptive Otsu based thresholding
Revathy Sivanandan, J. Jayakumari · Research on Biomedical Engineering · 2021
Purpose The advantages of ultrasound imaging are compromised by the presence of speckle, low contrast and intensity inhomogeneities especially in liver tissues, where the tumour is similar to the surrounding liver parenchyma. Accurate identification of the tumour region is vital and delineating the region of interest irrespective of the echogenicity is necessary. Methods In this paper, an active contour based segmentation is implemented where the contour initialization is achieved using an adaptive Otsu based thresholding. Different to existing methods, the image is neutrosophically enhanced and further despeckled using shearlets during pre-processing. The pre-processed image is thresholded with respect to predicates derived from mean, texture, gradient, phase and phase gradient. The phase gradient, which extracts the gradient information from the phase of the tumour image, is a novel concept introduced in the proposed work. The mask derived by the proposed thresholding is used as an initial contour for active contour segmentation based on local mean energies. This method of thresholding gave better mask when compared with simple Otsu thresholding or entropy thresholding. Results The proposed thresholding gave 5.02%, 33.91%, 34.02% and 15.65% increase and 90.7% decrease in structural similarity, Jaccard index, Dice coefficient, accuracy and mean square error respectively when compared with simple Otsu thresholding. Applying local statistics in active contours yield better results than global region based Chan-Vese segmentation. Localized energies based segmentation gave 10% more accuracy and 37.18% less error than global energy based segmentation. Conclusion The proposed method was applied on ultrasound liver tumour images including challenging isoechoic, slightly hypoechoic and heavily shadowed echogenecities, and showed promising segmentation results.