Professional Documents
Culture Documents
A Dual Channel Multiscale Convolution U-Net Methodfor Liver Tumor Segmentation From Abdomen CT Images
A Dual Channel Multiscale Convolution U-Net Methodfor Liver Tumor Segmentation From Abdomen CT Images
Abstract- Nowadays, the automatic segmentation of liver tumor appropriate treatment options and help clinicians diagnose
is needed for various medical applications like pathological cancer with a quick assessment of surgical treatment [2].
detection of liver diseases, surgical planning, and postoperative The Liver Tu mor Seg mentation (LiTS) challenge has been
evaluation. The liver is a massive, meaty organ on the right
structured for automatic and manual methods since 2008
part of the abdomen that plays an important part in our
digestive system. Cancerous tumors in the liver can cause a
[3]. Several ensemble seg mentation algorithms, such as
serious threat to human life. Currently, automatic liver tumor region growing, adaptive thresholding, level-set, and
segmentation has a lot of problems, such as complex tissues, AdaBoost [4-5], were applied as the best method for
fusion of features of the same scale, insufficient learning of automatic segmentation. However, the variability of liver
features, and irregular shapes of tumors. Typically, more lesions and the appearance of darker and brighter
networks are used to learn the important features, and fuse the surroundings, makes it difficult to design an automatic
features of same scale between encoders and decoders. But still, segmentation task. In vision-based computer tasks, Deep
segmentation of tumor is one of the main challenges in the Learn ing (DL) techniques have made significant
liver. Besides, soft organ removal, noise removal, and edge
contributionsin imp roving the efficiency of segmentation.
detection are important stages in the segmentation of liver
tumor. Aiming at these problems, this paper introduced a Dual
The neural network model ensembles the U-net model to
Channel Multiscale Convolution Unet (DCMC-Unet) to learn features automatically for liver seg mentation and
segment the liver tumor. An Intensity based threshold method refine them within the liver bounding box fro m the acquired
(IBTM) is used for removing the soft organs, and Z-score data [6]. A mong these facts, segmentation techniques must
normalization is utilized to eliminate the noise. The Enhanced deal with several difficu lt tasks, such as blurring tumo r
sobel method (ES M) is used to enhance the local features of the boundaries, and varying volu me and shape of the tumo r
liver. Finally, the enhanced features are fed to the DCMC- fro m patient to patient [7-9]. Therefore, further
Unetmodel to segment the tumor from the liver. Experimental improvement in auto mated methods is needed to accurately
analysis illustrates that the DCMC-Unet approach attains
segment liver tumors on CT images.
higher segmentation results than existing models in terms of
several qualitative metrics.
In this paper, we have introduced the DCMC-Unet method
Keywords: - Computed Tomography (CT), Deep Learning (DL), to accurately segment the liver tu mor. Here, we utilized
Liver Tumor Segmentation, U-Net, Liver Tumor Segmentation fusion of mult iscale features in the similar coding layer,
Challenge (LiTS). which increases the accuracy and can be used to learn more
information from the CT images .
I. INTRODUCTION
The second leading cause of death in recent years is the The contributions of the paper are given below:
Liver cancer.There will be appro ximately 830,000 deaths
world wide fro m liver cancer by 2020, an estimate report A DCMC-Unet is utilized to join the different scale’s
fro m World Health Organization (WHO). An effective features, so more information regarding tumor are
diagnosis method is needed to provide accurate information learned.
about the liver tumor during its earlier stages. Presently, the It also avoids feature data loss because of the fusion of
most popular imaging method for d iagnosing the liver features between the encoding and decoding layers.
cancer is Co mputed Tomography (CT) [1]. This can provide
The CT images of the liver have variance in the Hounsfield have numerous nesting and skip connections between
units, which leads to the high fluctuation in the intensity of decoding and encoding layers. This results in the loss of
the image. Ho wever, there is a large amount of noise in the some features and the learning of insufficient info rmation
images obtained from any clin ic or hospital, which must be fro m mu ltiple scales. Therefore, we present DCM C-Unet.
reduced before the segmentation stage. To overcome this The U-Net convolution module is rep laced in the coding
problem, a normalization method is necessary to ensure that layer by our multiscale dual-channel convolution module.
all non-zero pixels in the image have a unit variance and Initially, the image is led to the network. A dual channel
zero mean. split convolution process takes place after the operation of
dual convolutions, and the image is enlarged using a
B. Edge Detection Stage
similar up samp ling operation, therefore the network may
learn mo re features. Next, by using max-pooling and dual
The de-noised image is fed to the edge to enhance the local
convolutions operations, the image is en larged to the
features by using ESM. The Sobel filter's main purpose is to
same input image size, mu lti-scaling is done between the
detect edges by using the Sobel operator. It generates an
input image and the dual channel split convolution feature
image's gradient in both horizontal and vert ical dimensions
fusion, and a residual connections are added, which is
using convolutional templates.
shown in fig.2. Therefore, the DCM C-Unet can learn the
An illustration example of Sobel significant details of tumor by feature fusion.
convolution kernels is shown in Figure 2.
256×256×64 256×256×64
128×128×128 128×128×128
64×64×256
Figure 2. Sobel convolution kernels 64×64×256
B. Evaluati on Metrics
A. Evaluati on Database