Benign Hepatic Tumor Segmentation on Ultrasound Images

Abstract

Benign hepatic tumors are although noncancerous in nature, but if they are not monitored regularly they might cause problems in their latter stages. There are two categories of benign hepatic tumors, first category can be defined as semi-solid masses i.e. sacs like structure filled with fluid and the second category contains tumors that are solid masses. These lesions although do not spread to other parts around the affected area but if the size keeps on increasing it might damage the liver severely. Tumor segmentation is one of the most common non-invasive methods that is used to monitor the tumor size efficiently. Segmentation of hepatic tumor on ultrasound (US) images is a complex task due to intratumoral intensity inhomogeneity and similarity in the tumor appearance with rest of the liver region. Level set method (LSM) is one of the segmentation methods that is capable of performing segmentation on ultrasound images and provide accurate results. This segmentation method is mostly preferred as the curve can split and merge to take the topology of the desired tumor. A higher order function which is usually a distance function is used as a level set function. When this level set function is at zero level, it is marked as the initial curve. Then by using different parameters, image based energy is formulated such that this energy will have minimum value at the tumor boundary. To guide the curve evolution from its initial location towards the tumor boundary efficiently, a regularizing term is added with the image based energy. Based upon different characteristics of solid and semi-solid type of hepatic tumor on US images, two different methods were developed to segment them using LSM. Chan and Vese used Mumford-Shah fitting term by considering intensity variations to develop the image based energy whereas Georgiou et al. used distance between the intensity distribution inside and outside the curve as a parameter to segment the desired region. The drawback of these methods was that they were not very accurate on images with intratumoral inhomogeneity and suffers from curve leakage at weak edges at the boundary of the tumor. In the present work, some modifications are introduced in these methods. A shape based stopping criteria is introduced in the method developed by Georgiou et al. to avoid leakage problem. The regularization term of both methods are modified to guide the evolving contour towards the exact boundary of the tumor. The result of modifications introduced clearly demonstrates that they outperform the original methods. The proposed method is also compared with existing segmenting methods in terms of disc similarity coefficient and relative volume difference to show the high accuracy and reliability of the present work.