Aims

The liver CT image segmentation is still until now a challenging problem due to the fuzzy nature of the tumor transition to the surrounding tissues. The objective of this article is the consideration of the uncertainty present around the boundaries of a tumor region in the segmentation of the liver CT images as well as the segmentation of multiple tumors in the same CT image.

Materials and methods

A semi-automated segmentation method, including entropy-based processing, is proposed to segment single and multiple liver lesions from CT images. The proposed method introduces an entropy-based fuzzy region growing (EFRG) technique to extract the liver tumors whilst reducing the leakage, notably in CT images including several lesions. In fact, after the image rehaussement, an entropy-based fuzzy region growing was introduced in order to take into account the fuzzy nature of the tumor transition to the surrounding tissues. In fact, the local entropy is computed for each pixel in order to consider the spatial distribution of gray levels and to represent the variance of the local region. Then, after selecting manually the seed pixel, fuzzy membership function is used to preserve the fuzzy nature of tumor boundaries and to postpone the crisp decision until further information can be available to make the final decision. Starting with the seed pixel, the proposed method iteratively computes the region mean entropy and the resulted tumor region is obtained using a fixed threshold-based membership degree.

In the case of multiple tumors in the same liver CT image, the overlapping between adjacent tumors is treated through a distance-based processing in order to assign each pixel exclusively to one tumor.

Results

Experimental results prove that the method accurately segments single and multiple tumors in liver CT images over three different datasets, despite their small size, heterogeneity and fuzzy boundaries. Results were evaluated using standard quantitative measures, including the area overlapping error (AOE) and the relative area difference (RAD), for 2D segmentation, the volume overlapping error (VOE) and the relative volume difference (RVD), for 3D segmentation, and Dice similarity measure (DSM) for both cases. The mean AOE, RAD and DSM values reached by the entropy-based fuzzy region growing method over the ImageCLEF dataset were 19.9, 15.45 and 0.88, respectively. The results show that the proposed method is equivalent or even better to state-of-the-art methods over the 2D NUH dataset as well as over the 3D MIDAS dataset.

Conclusion

An entropy-based fuzzy region growing method was proposed to treat the overlap between overlapped tumors in liver CT images. This allows to improve results compared to the liver CT image segmentation methods of the state-of-the-art.