Automated Classification of Liver Cancer Stages Using Deep Learning on Histopathological Images

Liver cancer carriages a significant menace to health of the human and life expectancy. Hepatocellular carcinoma (HCC), a main liver cancer originating after liver cells, is an aggressive form of cancer. Each year, approximately 905,677 original cases of liver cancer are identified globally, with an impermanence rate as high as 90 percent. Liver cancer is the third important cause of cancer death globally. In 2020 alone, over 747,000 deaths from liver cancer were reported across the globe, in addition to more than 830,120 new cases. So, China accounts for around half of the universal problem of liver cancer. Being able to identify early what stage the liver cancer is at is really important for deciding the best treatment plan going forward. The potential of these advanced algorithms for machine learning to identify malignant liver lesions at an early stage is remarkable. For people who have this kind of cancer, more rapid and successful medical therapy may be possible due to the increased detection. Overall, developing liver cancer detection tests as well as therapy options through ongoing research remains a critical globally health concern.

Understanding precise staging for cancer of the liver is crucial for creating customized therapies for each patient. The accurate data that these deep learning algorithms may offer on the amount and spread of the disease allows for the development of customized therapies. By contrast, typical microscope assessment of samples of tissue might be personal, with varying views across clinicians about what they observed [1, 2]. The use of deep learning ushers in a new era of consistent and impartial analysis, minimizing differences in medical professionals' viewpoints and guaranteeing precise results.

Correctly identifying the stage of cancer of the liver may significantly enhance the chances of receiving optimal treatment. Doctors can employ special medications, radiation therapy, or an operation, according to the stage. Using deep learning for liver cancer staging has sped up research in medical imaging and pathology [3, 4]. This has helped doctors understand the disease better and develop modern ways to diagnose and treat it.

Catching liver cancer early and staging it precisely can save a lot of money for our healthcare system. That's because if caught late, treatments are more advanced and expensive. As deep learning keeps getting better, it will help create even more advanced models for diagnosing liver cancer stages. That will help doctors manage the disease even more.

1.1 Limitations of current diagnostic methods

Liver cancer has emerged as a formidable threat to human life and health over recent years, with both its global incidence and fatality rates on a steady incline [5-7]. The primary diagnostic methods currently employed comprise image tests, serum tests, and biopsies. Histopathological scans provide a detailed portrayal of liver cancer features, such as satellite nodules, metastases, and adjacent lesions, as well as crucial factors like the degree of differentiation, size, location, number, type, cellular, capsular, and vascular invasion.

Among these factors, the degree of differentiation significantly influences the malignancy level. The malignancy decreases as the cells become more differentiated from each other and from the normal tissue cells. Conversely, the malignancy escalates as differentiation decreases [8, 9].

Given the importance of diagnosing liver cancer with diverse degrees of differentiation, precise classification of histological images of liver cancer is both essential and irreplaceable. However, classifying histopathological images with different levels of differentiation presents challenges. It is not only time-consuming and labor-intensive, requiring substantial manual effort, but it is also prone to errors due to the subjectivity of the individual and the variable experience levels of physicians. Such misjudgements can significantly impact the formation of a patient's prognosis and treatment plan [10].

Hence, the ongoing research into the classification of liver cancer histological images carries substantial relevance and promise [11, 12].

1.2 Role of AI in medicine

In the wake of rapid advancements in artificial intelligence, the medical domain has extensively leveraged AI algorithms in recent years. Deep learning, an AI technique predicated on deep neural networks, constitutes one such artificial intelligence algorithm [13]. Computer vision and NLP tasks frequently employ deep learning, especially in the arenas of early disease diagnosis and medical image processing [14, 15].

In a significant study, Krishan and Mittal [16] utilized six distinct classifiers to categorize various tumor stages derived from CT scans. Their tumor classification accuracy spanned from 77.47% to 89.12%, thereby enabling radiologists to expedite the diagnosis and treatment of liver ailments substantially. However, this study also has limitations such as the exclusive use of one type of sample.

Xu et al. [17] proposed a radiomic diagnosis model that effectively distinguished between intrahepatic cholangiocarcinoma (ICCA) and hepatocellular carcinoma (HCC) using CT images. With AUCs (Area Under the Curve) of 0.847 in the assessment cohorts and 0.659 in the validation, this model will facilitate differentiation between ICCA and HCC in future. Nonetheless, this study underscores the need for enhanced classification accuracy.

Wan et al. [18] presented the MMF-CNN, a novel methodology that integrates multi-level and multi-scale fusion techniques within Convolutional Neural Networks (CNNs) for the purpose of detecting liver lesions in magnetic resonance imaging (MRI). They systematically used their inventive technique on several cutting-edge deep learning systems. The results of the research support the efficacy of their proposed strategy and highlight the way it may enhance the accuracy of MRI images used for assessing lesions in the liver.

Zhou et al.'s [19] comprehensive examination of machine learning (ML) and deep learning (DL) methods established a foundation for scientific literacy in the artificial intelligence (AI) field. In addition, they concentrate on convolution neural networks (CNNs) and its application to liver-related imaging issues. The real-world study in the article shows how AI can be used to detect and assess specific liver areas, enhance the results of therapy, and predict the liver's reactions to medicines. Their cooperative study supports machine-assisted health as a possible development in liver-related therapies.

In the same vein, Sureshkumar et al. [20] developed an entirely automatic Computer-Aided Diagnostic (CAD) approach for detecting hepatocellular carcinoma (HCC) employing CNN design. The present investigation has investigated a variety of techniques and algorithms used to assess cancer of the liver.

The suggested approach seeks to enhance overall precision while identifying rare tumours with a smaller rate of false positives. In addition, the study provides an original method to categorize liver scans with CT and determine between regular and unusual patterns. When put together, these findings present exciting novel perspectives as to how AI could enhance liver disease diagnosis and paving the door for better and more precise treatments.

1.3 Deep learning in liver cancer diagnosis

Due to the limitations caused forth by fewer participants, the study left the diagnosis criteria unchanged. In Lin et al.'s study [21], a Convolutional Neural Network (CNN) was carefully constructed utilizing the VGG framework. This contributed to a classification accuracy of more than 92% for all stages of liver cancer.

Using a deep learning algorithm together with multi-photon imaging demonstrated potential in differentiating between various forms of hepatocellular carcinoma (HCC), opening up fresh possibilities for automated testing [22]. There is plenty of potential for enhancement of accuracy, given that the study is still in its infancy. This work enhances the process of effectively collecting characteristics from medical visuals by incorporating a visual focus system to the deep learning framework.

Figure 1 shows the architecture diagram. Initially input image is taken from the Kaggle Dataset and pre-processed with Weiner filter then segmented and extracted features using FCM-CSA and SIFT algorithms respectively. Extracted features are classified using R-CNN and liver tumor is classified with different classes. This classification helps the medical experts to take the right decisions at right time to save the human life. Gray-Level Co-occurrence Matrix (GLCM) can be a valuable tool for feature extraction in liver cancer identification, especially when analyzing medical images like CT scans, MRIs, or ultrasound images for texture analysis, discriminative features, local analysis and interpretability. For liver cancer identification need to features are very important for classification so only GLCM was utilized in the proposed work.

qqjie_tu_20240229134622.png

Figure 1. Architecture diagram for proposed method

R-CNN (Region-based CNN) is powerful computer vision techniques that have been widely used in various applications, including object detection, image segmentation, and more. While these techniques are not specific to liver cancer identification and classification, they can be essential tools in the field of medical image analysis, including liver cancer identification and classification, for several reasons like accurate detection, localization, feature classification, and extraction.

While R-CNN and its variants offers many advantages, it's significant to note the success be contingent on the availability of high-quality annotated medical image datasets and the expertise of healthcare professionals to interpret the results and make informed decisions based on the model's predictions.

3.1 Image pre-processing

The foundational step is pre-processing, a crucial stage in which input images are sourced from both the Kaggle dataset and on-site field visits. During pre-processing, the objective is to eliminate undesirable noise from the images while simultaneously boosting the contrast levels to enhance overall image quality.

In liver cancer identification and classification, pre-processing prepares the raw medical images for subsequent feature extraction and AI based algorithms. It ensures that the data used for analysis is accurate, consistent, and conducive to reliable results, ultimately improving the diagnostic and classification accuracy for liver cancer.

Figure 2 (a) shows input image of Histopathological image of liver with Cancer and Figure 2 (b) shows the pre-processed image of liver with cancer cells.

3.2 Segmentation using Fuzzy C-Means (FCM)

In the segmentation of diseased (cancerous) portions, an innovative approach called FCM-CSA (Fuzzy C-Means based Chameleon Swarm Algorithm) was applied. The segmentation process within FCM-CSA is delineated through the following steps: Leveraging the principles of fuzzy logic, an unsupervised classification model akin to the Fuzzy C-Means (FCM) network was introduced. Unlike traditional partitioning models such as k-means, where each data point corresponds to a single partition, the FCM model allows for a more nuanced representation of data points within multiple partitions, providing a richer and more flexible segmentation framework.

qqjie_tu_20240229134910.png

(a)

qqjie_tu_20240229140332.png

(b)

Figure 2. a) Histopathological image of liver with cancer b) Pre-processed liver with cancer

Eq. (1) calculates the cluster centers:

$D j=\frac{\sum_{K=1}^M V_{j, k Y_k}^n}{\sum_{K=1}^M V_{j, k}^n}$            (1)

The dataset comprises M data points, denoted as y_k, where k represents the index of each pixel. The degree of membership for the kth pixel is expressed through a fuzzifier denoted as 'n,' with the constraint that 'n' is greater than 1. Eq. (2) articulates the formula that encapsulates this relationship, providing a mathematical representation of the degree of membership for each pixel in the dataset.

$V_{j, k}=\frac{1}{\sum_{L=1}^{m d} \frac{C_{j, k}^{2 /(n-1)}}{C_{j, k}}}$            (2)

Degree of membership is represented by v(j,k). c2(yk, dj) gives the distance that exists between the jth group and the kth pixel. The parameter 'n', which provides a fuzziness level where 1 < n ≤ ∞, plays an essential part when determining the number of membership degrees. The method described in look at 1 yields ideal clustering centers, which can be utilized to help segment the malignant disease off the image.

The Fuzzy C-Means (FCM) method has certain issues throughout the segmentation process, particularly with the initial cluster centroids' poor performance. In order to accomplish this, we employ a mixed approach that involves the FCM-based Chameleon Swarm Algorithm (CSA), additionally referred to as FCMCSA. The meta-heuristic technique known as the CSA begins the optimization procedure with initializing the number of participants. The first population is formed with a random initialization in a search space, assuming an overall population count of 'C' within the search space of size 'D'. By utilizing the benefits of both FCM and CSA, this combination approach reduces the limitations of FCM's original cluster centroids and enhances segmentation quality overall.

a_i = L_j + rand × (U_j – L_j)              (3)

The first vector of i^th chameleon is represented by a_i, with the upper and lower limits of the search area defined as L_j and U_j in the jth dimension, respectively. The variable rand denotes a randomly generated number within the range of 0 to 1. The reinforced exploratory capabilities of the chameleons in the search space are succinctly articulated as follows:

$\rho=\delta \exp ^{\left(-\frac{\alpha t}{R}\right)}$            (4)

In the iterative process, the parameter ρ plays a pivotal role and diminishes as the iterations progress. Predefined parameters δ, α, and β are strategically employed to govern the delicate balance between exploration and exploitation stages. The updating of chameleons' positions in the search space is orchestrated through rotating centered coordinates, coupled with an acceleration factor. With Figure 3 offering a representation of the input image alongside the segmented output achieved through the FCM-CSA.

qqjie_tu_20240229134622.png

(a)

qqjie_tu_20240229134910.png

(b)

Figure 3. a. Binary image; b. Segmented image using FCM-CSA algorithm

3.3 Feature extraction using GLCM

Following segmentation, the next step involves feature extraction through a rapid GLCM (Gray Level Co-occurrence Matrix) feature extraction model. Typically, GLCM features provide comprehensive texture descriptions of images, but their computational intensity poses a significant resource challenge. In this model, the gray levels of image pixels undergo re-quantization to minimize G, enhancing the efficiency of the GLCM feature extraction process. The fast GLCM model proves to be a robust method for extracting features related to liver cancer disease. The model operates by establishing connections between spatially close pixel features, calculating GLCM matrices under the correlation type, and assigning index pixels to all pyramid weight matrix neighborhoods. To estimate GLCM features for other pixels, this model, as illustrated in Eq. (5), is employed, ensuring a comprehensive and efficient extraction process.

$F(q)=\sum_{j=1}^m b j(q) F(y i)$            (5)

The feature vector for non-key pixels, denoted as F(q), is defined, with each overlapping weight window centered on a key pixel equal to y_j. To assess the efficacy of the fast GLCM performance, simulations are conducted with varied image sizes, utilizing three single-band images for validation. The simulation employs the following parameters, and the results are visually depicted in Figure 4, showcasing the extracted features from a histopathological image depicting liver cancer using GLCM.

qqjie_tu_20240229134622.png

Figure 4. Feature extraction using GLCM

• Utilizing a fixed window size of 33 × 33 for GLCM extraction simplifies the process. The extracted features using the GLCM algorithm are illustrated in Figure 4.
• The execution load of GLCM is effectively reduced, and sparsity of GLCM matrices is reduced through strategic adjustments.
• To diminish the computational load (G), a re-quantization of the input image gray levels is performed.

3.4 Feature classification using fast R-CNN

A powerful object detection tool, known as the Fast Region-Based CNN (Fast R-CNN), employs deep learning techniques for categorizing objects within an image. Leveraging the Edge Boxes approach, it enhances the computational efficiency of object detection by generating effective region proposals. Edge Boxes provide a straightforward measure of objectness, subtracting edges that are part of contours crossing the border of the box's.

Possible object-containing areas are initially meticulously delineated using Edge Boxes in the two-stage Fast R-CNN recognition technique. In the next phase, each component is classified. In the first step, layers of convolution are employed to analyze the entire picture with the goal to extract features. In addition to generate region ideas and extract characteristics from the image, Edge Box are additionally processed at this stage.

Regions of interest (ROIs) in hepatic pictures are identified and categorized via Fast R-CNN, which finds application to liver cancer classification. The assessment of Fast R-CNN's performance in detecting liver cancer can be performed using a number of criteria, such as precision, recall, precision, accuracy, and F1-Score. The device's operating characteristic curve's area under the curve (AUC-ROC).

Accuracy calculates the precision of positive diagnoses among every sample marked as positive, while accuracy estimates the proportion of properly classified samples. Among every one of the genuine positive samples, recall assesses the precision with which positive classification have been generated. Recall and accuracy are evaluated in an equitable way by the F1 score. The ability of the model to differentiate between samples that are positive or negative across a range of threshold is assessed with AUC-ROC.

A set of annotated liver images that can tell the difference among malignant or non-cancerous material is essential to evaluate the efficacy of Fast R-CNN in cancer of the liver diagnosis. Three sets of data points have been divided into test, validation, and learning. The model changes on its initial set throughout training, changing hyperparameters for learning rate and batch count for optimal. These evaluations help identify the model's benefits and drawbacks, directing potential enhancements to the layout and training methods. Three different categories utilizing the Fast R-CNN technique are shown visually in Figure 5.

qqjie_tu_20240229134622.png

Figure 5. Different classes of classification using R-CNN algorithm

3.4.1 Training

By utilizing an image the labeler is an essential for object detectors training such as Fast R-CNN. The image labeler program facilitates the assignment of predetermined rectangular Regions of Interest (ROIs) labels to a set of photographs, distinguishing them into three types: poorly differentiated, well differentiated, and moderately differentiated. During the training process, expert doctors are actively involved in classifying liver cancer from low to high grades.

The labeled ROIs for each image can be exported to the MATLAB workspace. After creating ROIs and labels for each image in a collection, a command like "trainingData" can be used. This exported ground truth serves as crucial input during the training phase.

The object detector introduced in this model; Fast R-CNN is built upon the foundation of the VGG-16 Convolutional Neural Network (ConvNet). The VGG-16 ConvNet undergoes a transfer learning approach, adapting pre-trained weights for effective object detection. Each image requires a corresponding label identifying the rectangle region of interest (ROI). In this study, the region for each image in its label column is set to [1, 1, 224, 224], given that every image in the dataset is a cropped image of liver cancer.

Modifications to the VGG-16 model include adjusting the dimensions of the last three layers and the image input layer from [224, 224, 3] to [48, 48, 3]. Additionally, the grayscale picture dataset undergoes transformation from [48, 48, 1] to [48, 48, 3]. Further enhancements involve the addition of a region-of-interest pooling layer and a bounding box regression layer, transforming the pre-trained model into Fast R-CNN object identification network. This adaptation ensures the effective integration of the VGG-16 model into the Fast R-CNN framework for accurate liver cancer classification.