ESP-UNet: Encoder-Decoder Convolutional Neural Network with Edge-Enhanced Features for Liver Segmentation

Cancer is the second most common cause of death worldwide, according to the WHO. The World Health Organization (WHO) reports that 8.8 million people died from it in 2015, with liver cancer accounting for 788,000 of those deaths [1]. The liver is the body's biggest gland. The liver digests, detoxifies, and metabolisms. Liver cancers are classified as primary or secondary. Primary liver cancer begins in liver tissue. Primary liver cancer includes hemangioma and hepatocellular carcinoma (HCC). HCC is the most prevalent liver cancer caused by liver tissue malignant cell development. Blood vessels form a hepatic hemangioma [2]. Secondary metastatic liver cancer occurs when cancer spreads from another bodily organ [3]. The liver's texture and shape alter due to an abnormality. Liver, vascular, and tumor excision and segmentation are needed for illness diagnosis. Due to liver homogeneity, poor contrast, form, and adjoining abdominal organs, proper liver segmentation is difficult. Medical imaging methods including CT, MRI, and US may identify liver disorders [4]. The last decade saw the adoption of statistical, threshold-based, fuzzy logic-based, clustering, artificial neural network (ANN) models, and machine learning-based (ML) ALS schemes [5]. Hand-crafted features have lesser interconnectivity, feature distinctiveness, and correlation to original image features. The conventional ML classifier’s effectiveness is highly reliant on these features which quality and distinctiveness can be easily degraded by noise, blur, illumination variance, etc.

This paper presents liver segmentation using edge-strengthening parallel UNet to improve the segmentation performance of the boundary area of the liver. The chief contributions of this article are summarized as follows:

• Encoder-Decoder Convolutional Neural Network (ESP-UNet) with edge-enhanced features for liver segmentation to tackle the problem of under-segmentation and over-segmentation.
• Performance estimation of ESP-UNet on LiTS dataset based on accuracy, dice score coefficient (DSC), Volume Overlapping Error (VOE), Jaccard Index (JI), and Relative Volume Difference (RVD).

In recent years, deep learning (DL) becomes extremely popular for medical image segmentation because of its ability to deal with complex segmentation problems, accurate segmentation, ability to work for the datasets, etc. Vadali et al. [6] suggested a model for analyzing liver cancer using a support vector machine (SVM) algorithm in MATLAB. The proposed model works in four stages, image resizing, segmentation using stage level set scheme to remove an unwanted area, post-pra processing, and classification stage. In the last stage, the VM is used for the classification of images into benign and malignant. The proposed model achieved a Precision rate of 86.7%. Mouhni et al. [7] highlight the recent surge in using Deep Neural Network algorithms, particularly CNNs, for medical imaging. They point out the challenge of acquiring large datasets, especially for rare diseases. To address this, they introduced Federated Learning (FL) as a potential solution, allowing multicenter dataset training without breaching data privacy. The study offers a comprehensive review of FL in medical image analysis and its associated strategies. Later, Dutta and Dubey [8] proposed a model based on MRI images. The watershed technique was utilized to separate cancer cells from an MRI scan image, and Otsu's method was used to improve the image. The wavelet transform method is used to extract features. The wavelet transform has a multi-resolution analytic that can analyze pictures with varied resolution units. In 2020, Dong et al. [9] proposed a model using Hybridized Fully Convolutional Neural Network (HFCNN) for liver segmentation and tumor detection. The model involves training various layers of HFCNN. The system uses a 2D CT image as input. The proposed framework has shown a high accuracy of 97.22% for liver volume measurements and a 0.92 average dice coefficient. Abdulgani and Al Ahmad [10] anticipated an unsupervised method for cancer detection. Here classification using optical measurement is utilized that interact differently with light as various cells has distinct composition. In the proposed work optical measurement is combined with complex schemes for the classification of cancer cells. In this approach, six different kinds of cells including cancer cells were assessed based on transmission characteristics. The optical response of individual cells was modeled with some parameters such as damping factor, frequency, amplitude, etc. The parameters are extracted using Prony method. Further, Xi et al. [11] proposed deep patch CNN for segmentation and detecting High-resolution cancer. High-resolution medical images cannot be directly given as input to neural networks, because the large size of these images will require very high computing ability and large training time. Therefore, medical images are first compressed and then given as input to the neural networks. Compressing medical images causes the loss of some critical information that may be essential for detecting any anomalies in medical images. The author offers a novel method for building abnormality detectors to address this issue, which involves integrating pre-trained deep CNNs with class activation mappings and area proposal networks after optimizing them on image patches focused on clinical abnormalities. On a mammography dataset, the deep patch classifier produced an average classification accuracy of 92.53% as opposed to 81.55% for a conventional method based on manually created features. On an ultrasound liver imaging data set, the integrated detector was evaluated for anomaly detection and received a mean precision of 0.60. Later, Xi et al. [12] proposed a novel microcantilever biosensor for detecting liver cancer. Cantilever sensors, out of all the biomarker sensors, are the most appealing choice for real-world applications because of their low cost, high sensitivity, quick response, and mobility. Static and dynamic modes are available for cantilever-based sensors. The mass loading causes a shift in resonance frequency, which the lever senses. Cantilever is capable of measuring sensitivity in the fem to gram to atto gram range. Subsequently, Sekaran et al. [13] proposed a framework that uses a DL strategy named CNN, embedded with Gaussian Mixtmaximizationd and Expectation maximization (EM) algorithm to predict pancreatic cancer from liver CT images. This model was able to predict the cancer spread in the pancreas. The experiment was carried out using a cancer imaging archive (TCIA) which consists of approximately 19000 images. The project was named Fishman Afterward with high accuracy. Afterward, Khan and Loganathan [14] proposed a system (AutoLiv) for the automatic segmentation of liver tumors. The authors utilized CT images as the input to the system and then applied a threshold-based slope differentiation technique [SDD] for segmentation. Upon the segmented image, alternative fuzzy c-means clustering [AFCM] is applied to detect the tumor. The proposed system is implemented using MATLAB and then compared with the existing benchmarks. Fan et al. [15] investigated multi-scale context nested UNet (MSN-Net) for ALS which helped to decrease the semantic gap and gradient vanishing problem. It decomposes the low and high levels to enhance the distinctiveness of the network. Bai et al. [16] anticipated multi-scale candidate generation (MCG), active contour model (ACM) and 3D fractal residual network (3D FRN) for ALS to tackle the problem of location, size and shape variation of liver tumor. It outperformed the traditional ALS approaches but was restricted due to poor performance for borderline tumor detection and multiple tumor detection. Li et al. [17] explored hybrid densely linked UNet (H-DenseUNet) for ALS that resulted in a DSC of 0.93. The H-DenseUNet structure is more complex and needs higher trainable parameters because of the use of parallel training.

From the review of various recent liver segmentation techniques, it is observed that deep learning-based approaches outperform conventional region-based and intensity-based methods. However, the results of these methods are greatly dependent on structure, contrast, noise, borders, shape, and imaging technique of liver images as very few samples are generally available for training. The inaccurate border segmentation of the liver leads to an inappropriate diagnosis of liver disease and imprecise localization of the liver. The conventional deep learning techniques provide inferior results for the less dataset. Very less concentration is given to the border segmentation of the liver which plays an important role in the location of the liver in abdominal CT images. Thus, there is a need to provide effective liver segmentation which can deal with non-uniform illumination conditions, piercing borders, under-segmentation, and over-segmentation.

The rest of the article is arranged as follows: Section 2 provides a brief discussion of the proposed methodology of liver segmentation. Further, Section 3 elaborates on the experimental results and discussions. Lastly, Section 4 provides a succinct conclusion and offers direction for future enhancement.

The suggested ESP-UNet is simulated using Python 3.8 and the OpenCV toolbox on a Nvidia GPU having 512 Tensor-core, 16 GB RAM, and 32 GB graphics, and a Windows operating system. The proposed method is validated using the LiTS dataset [23], which includes 131 CT volumes with varied tissue abnormalities, contrast liver size and tumor levels. The performance of the proposed algorithm is evaluated using a binary mask of segmented liver (LS) image and ground truth (GT) images based on accuracy, DSC, VOE, JI and RVD. The DSC provides information regarding complete ALS from abdominal images that range between 0 (no overlap) and 1 (complete overlap). Eq. (3) is used to compute the DSC.

$D S C(L S, G T)=\frac{2 \times(L S \cap G T)}{|L S|+|G T|}$       (3)

The RVD signifies the changes that occurred between the segmented liver image and GT that lies between 0 (Good segmentation) and 1 (worst segmentation). The RVD is computed using Eq. (4).

$R V D(L S, G T)=\frac{|G T|-|L S|}{|L S|}$       (4)

The VOE represents an overlapping error between the segmented liver image and GT that lies between 0 to 1. Zero value represents better segmentation. The VOE is computed using Eq. (5).

$V O E(L S, G T)=1-\frac{2 \times(L S \cap G T)}{|L S|+|G T|}$       (5)

The JI measures the similarity between ground truth and segmented image. The JI is computed using Eq. (6).

$J I(L S, G T)=\frac{D S}{2-D S}$       (6)

The visualization of the results of the proposed ESP-UNet is illustrated in Figure 5. The original liver image is used for liver object segmentation and Kirsch’s gradient image obtained using eight directional filters is used for the borderline liver segmentation.

5.png

Figure 5. Visualization of results of proposed ESP-UNet segmentation a) Liver CT image b) Liver segmentation using UNet-1 c) GT for Liver d) Kirsch’s filter output e) Final segmented liver image

The segmentation outcomes of the ESP-Net for sample images from the LiTS dataset are illustrated in Figure 6. It is noted that the proposed ESP-UNet provides superior performance compared with MSN-Net [15] and provides sharper border segmentation compared with the previous approach.

6.png

Figure 6. Segmentation results of ESP-Net

The performance of the ESP-UNet technique is compared with the earlier traditional state of arts for ALS such as ACM-ResNet [16], H-DenseNet [17], and MSN-Net [15] as described in Table 3.

Table 3. Performance comparison of proposed ESP-UNet with previous techniques

The proposed ESP-UNet provides 1.76%, 2.34% and 43.13% improvement in DSC compared with recent MSN-Net, H-DenseUNet, and ACM-ResNet respectively based on liver segmentation on the LiTS dataset. The ESP-UNet provides 95.99% overlapping of the segmented liver and GT image. The suggested ESP-UNet improves the borderline segmentation of the liver that avoids the under-segmentation and over-segmentation of the liver. This results in the proposed edge strengthening based on Kirsch’s operator and UNet gives significant improvement in VOE (2.5%- 87.96%) over the traditional techniques. Therefore, the ESP-UNet provides a VOE of 0.039 which has shown noteworthy improvement over MSN-Net (0.04), H-DenseUNet (0.116), and ACM-ResNet (0.324) for the LiTS dataset. The ESP-UNet-based encoder-decoder network supports diminishing under and over-segmentation by limiting the segmentation with the help of border segmentation of liver objects using Kirsch’s filter and UNet. When the results of ESP-UNet are compared based on JI it provides an improvement of 3.48%, 4.65%, and 84.2% over MSN-Net, H-DEnseUNet, ACM-ResNet respectively. The ESP-UNet provides the RVD of 0.09 which shows better segmentation compared with as H-DenseUNet (-0.01) and ACM-ResNet (0.194). Thus, it observed that the proposed ESP-UNet segmentation scheme is capable of providing better segmentation in complex CT images that considers the border of the liver object to limit the under and over-segmentation problem in the traditional liver segmentation techniques. The proposed scheme needs a training time of 128min and a segmentation time of 1.8 min. However, the parallel structure of ESP-UNet increases the complexity and total trainable parameters (~30M) of the proposed framework. The larger trainable parameters decrease the implementation flexibility of the proposed architecture on standalone medical diagnosis devices with limited resources. The effectiveness of the ESP-UNet is estimated for the LiTS dataset having limited samples. However, the performance of the proposed ESP-UNet can be varied for the larger dataset.