Segmentation of liver tumors in multiphase computed tomography images using hybrid method

Abstract

The multiphase scan has enabled an improved detection of liver tumors. However, tumor regions and peripheral tissues are difficult to distinguish and delineate owing to their highly similar image features. Moreover, their characteristics vary significantly in different phases. This is challenging when using segmentation methods that are based on unique training models. Herein, a hybrid framework is proposed for liver tumor segmentation in multiphase images. We first develop a cascade region-based convolutional neural network with refined head to locate the tumors. Meanwhile, phase-sensitive noise filtering is introduced to refine the segmentation conducted by a level-set-based framework. This method is sensitive to the intensity contrast but not to the regions of interest, thereby affording better performance in delineating adjacent tumors. In our experiment, the average precision and recall rates are 76.8% and 84.4%, respectively. The intersection over union, true positive rate, and false positive rate are 72.7%, 76.2%, and 4.75%, respectively.

Introduction

According to data from the Global Cancer Statistics 2020, liver cancer ranks third among the causes of neoplasm-related deaths worldwide [1]. Medical-image-based technology provides a conventional and valid non-invasive procedure for liver cancer diagnosis. Computed tomography (CT) imaging is one of the most typically used modalities for this task. The detection of liver tumors from CT is essential step in the diagnosis. However, tumors appear similar to the surrounding tissues, which increases the difficulty of tumor detection. To enhance the appearance of different tissues in a CT image, a multiphase CT angiography technique is introduced to provide physicians with multi-degree information as well as tissue differences based on the temporal resolutions [2]. The imaging time of CT is sufficiently short, thereby allowing each of the phases to be captured at a high resolution within a few seconds. A series of phase signals, i.e., non-contrast-enhanced (NC), arterial (ART), portal venous (PV), and delay phase signals, are obtained successively after contrast injection [3]. These characteristic variations from multiphase CT images allow radiologists to distinguish liver tumors [4], [5], [6].

Although the appearance of tumors is enhanced in a multiphase CT image, when compared with CT images without contrast enhancement, the detection of liver tumor regions remains challenging because of the wide variety of tumors appearing in the images from different patients. In fact, the intensity, texture, shape, size, and other features of the tumors are similar to their surrounding tissues or other independent tumors.

Recently, deep learning techniques have been introduced to medical image analysis and have provided new solutions for the segmentation task of liver tumor regions in multiphase CT. Conze et al. used the random forest classifier and multiphase supervoxel-based features to segment hepatocellular carcinoma liver tumors [7]. The multiphase visual features extracted from registered multiphase CT images have been proven to improve the accuracy of the classification result. Sun et al. [8] considered distinctive pathological information from different phases. They proposed a multichannel fully convolutional network with separate fully convolutional network (FCN) channels for each phase individually. The features of liver tumors, which involve different phases, are extracted in the deeper layers to improve the accuracy of segmentation. The efficacy of information and the features pertaining to the liver tissue from different phases can be corroborated in the study by Ouhmich et al. [9], where a dimensional multiphase was implemented on NC, ART, and PV slices. Linked single-phase images were formed as a feature map of multiple dimensions. Meanwhile, multiphase fusion was applied. The outputs from the single phases were combined to obtain the final results. To mine complementary information across images of multiple phases, a modality-weighted U-Net [10] was proposed to combine features learned from different phases using dynamically weighted feature maps. Meanwhile, Xu et al. substantiated the efficiency of information extracted from multiple phases in a residual network [11]. They used a phase-attention residual network to exploit the ART phase and then extracted tumors from the PV phase. The phase attention modules were separated to obtain channel-wise and cross-phase interdependencies. In addition, to enhance sensitivity to the edges, a three-dimensional (3-D) boundary-enhanced loss was used in the training.

However, these methods present two disadvantages. First, because the tumors appear tuberous, multiple tumor regions are likely to aggregate and form a single region. When the gap between the tumor regions is extremely narrow, as shown in Fig. 1, the segmentation method may fail. Second, insufficient data quantity restricts the application of relatively deep learning methods in medical image processing, particularly in abdominal multiphase CT images. Herein, a hybrid method that incorporates a deep learning method is proposed to improve the accuracy of liver tumor segmentation. A cascade region-based convolutional neural network (R-CNN) with expanded head sections was introduced to detect tumor regions separately. Based on the regions, tumor boundaries were segmented using an improved level-set algorithm.