Robust extraction for low-contrast liver tumors using modified adaptive likelihood estimation

Liver tumor extraction is essential for liver ablation surgery planning and treatment. For accurate and robust tumor segmentation, we propose a semiautomatic method using adaptive likelihood classification with modified likelihood model.

First, a minimal ellipse (or quasi-ellipsoid) that encloses a liver tumor is generated for initialization. Then, a hybrid intensity likelihood modification based on nonparametric density estimation is proposed to enhance local likelihood contrast and reduce its inhomogeneity. A prior elliptical (or quasi-ellipsoid) shape constraint is directly integrated into the likelihood to further prevent leakage of the algorithm into adjacent tissues with similar intensity. Finally, an adaptive likelihood classification is proposed for accurate segmentation of tumors with low contrast, high noise or heterogeneous densities.

Experiments were performed on 3Dircadb and LiTS datasets. The average volumetric overlap errors of the 3Dircadb and LiTS datasets were 27.05 and 35.72%, respectively. The algorithm's robustness was validated by comparing results of 5 operators with multiple selections on different tumors.

The proposed method achieved good results in different tumors, even in low-contrast tumors with blurred boundaries. Reliable results can still be achieved over different initializations by different operators using the proposed method.