Description:FIELD OF THE INVENTION
The present invention relates to a system for hepatocellular carcinoma staging and diagnosis. More particularly, the system utilizes a modified VGG16 and U-Net architectures, Transfer learning approach based ensemble model, to extract both global and local features from MRI images for early detection and precise staging of liver tumors, and better patient outcomes in hepatocellular carcinoma (HCC) management.

BACKGROUND OF THE INVENTION
Hepatocellular Carcinoma (HCC) is the most common type of primary liver cancer which poses a significant threat to global health. It stands as the second leading cause of cancer-related deaths worldwide, highlighting the urgent need for precise diagnosis and effective treatment strategies. This burden is further amplified by HCC's prevalence, accounting for roughly 90% of all liver malignancies. The severity of HCC is underscored by its intricate pathophysiology and often asymptomatic early stages. Early detection becomes paramount due to the disease's rapid progression and propensity to metastasize. However, achieving accurate staging – a crucial step in determining treatment plans and maximizing patient outcomes – presents a complex challenge.

HCC's widespread occurrence translates into a substantial global health concern. Its prevalence and high mortality rate create a significant financial, logistical, and emotional burden on patients, families, and healthcare systems.

This burden is particularly acute in regions with a high prevalence of chronic viral hepatitis infections, a major risk factor for HCC development. The disease's devastating impact extends beyond individuals, affecting families and societies as a whole. Early identification and precise staging are crucial for improving patient outcomes and mitigating the immense healthcare costs associated with HCC management.

The intricate relationship between HCC and chronic liver diseases like hepatitis B and C underscores the importance of addressing this public health challenge. These underlying conditions contribute significantly to the global burden of HCC, necessitating extensive research and cutting-edge treatment modalities to lessen its impact.

Accurate staging of HCC presents a multitude of challenges due to the disease's complex biology, diverse clinical manifestations, and multifaceted nature. Variability in tumor size and distribution further complicates the process. Unlike many cancers, HCC often progresses in a non-linear fashion, frequently involving multifocal lesions throughout the liver. This heterogeneity poses a hurdle for established staging techniques designed for unifocal tumor models.

The intricate link between HCC and underlying liver diseases like cirrhosis adds another layer of complexity. Cirrhosis alters the tumor microenvironment and influences disease progression, making it difficult to accurately assess HCC using traditional staging methods. The limitations of current staging systems, often based on clinical and pathological criteria, become apparent in their inability to fully capture the nuances of HCC and hence its early detection.

Furthermore, the absence of symptoms in the early stages often leads to delayed diagnoses, hindering the effectiveness of available treatments. While invasive tissue sampling techniques can provide valuable data, they also come with the drawbacks of sampling bias and other associated risks.

Given the intricate relationship between HCC and liver health, accurate disease identification and staging are of utmost importance. Early detection and precise staging are essential for determining treatment plans and improving patient outcomes. The intricate structure of the liver, the variability in disease presentation, and the potential for multifocal tumors necessitate advanced diagnostic methods capable of distinguishing between different stages of the disease. Accurate staging empowers physicians to tailor treatment interventions, ultimately enhancing patient survival and quality of life.

The asymptomatic nature of HCC during early stages presents a particular challenge. The absence of readily apparent clinical symptoms often delays identification until the disease has progressed to a more advanced stage, compromising treatment efficacy and significantly impacting patient survival rates. Therefore, proactive screening and early diagnosis are crucial for detecting HCC in its early stages when curative treatments offer the greatest chance of success.

The landscape of medical diagnostics is undergoing a paradigm shift fueled by the convergence of cutting-edge medical imaging technologies and transformative deep learning methodologies. Advanced modalities like computed tomography (CT), magnetic resonance imaging (MRI), and positron emission tomography (PET) offer unprecedented access to intricate anatomical and pathological data.

However, unlocking the therapeutic potential hidden within this complex and diverse data landscape requires novel computational approaches. Deep learning, a subfield of artificial intelligence, is emerging as a powerful tool to extract clinically relevant insights from medical images, particularly in cases like HCC where traditional methods fall short.

Convolutional Neural Networks (CNNs), a type of deep learning architecture, hold immense promise for medical image analysis. These networks are designed to autonomously learn complex patterns and information directly from images, a capability that has been harnessed for various tasks including disease classification and tumor segmentation.

U-Net, a specialized CNN architecture specifically designed for medical image segmentation tasks, exemplifies this groundbreaking development. Its unique U-shaped structure allows for the simultaneous capture of both local and global features within an image. This is particularly advantageous in biomedical imaging, where precise delineation of subtle anatomical details is critical for accurate diagnosis.

The application of deep learning in HCC diagnosis holds immense potential to address the limitations of current methods. By leveraging the power of CNNs to analyze complex medical images, researchers are developing novel algorithms capable of:
 Early HCC Detection: Deep learning models can potentially identify subtle changes in liver tissue indicative of early-stage HCC, enabling earlier intervention and improved patient outcomes.
 Improved Tumor Segmentation: Precise segmentation of HCC tumors is crucial for accurate staging and treatment planning. Deep learning algorithms can learn to distinguish between cancerous and healthy tissue with high accuracy, leading to more effective treatment strategies.
 Personalized Treatment Planning: By analyzing a patient's specific medical images, deep learning models can assist physicians in tailoring treatment plans to the unique characteristics of their disease.

The integration of deep learning into HCC diagnosis represents a significant leap forward in the fight against this global health threat. This technology offers the potential for earlier detection, improved staging accuracy, and ultimately, better patient outcomes. As research in this field continues to advance, a future where deep learning plays a pivotal role in transforming HCC diagnosis and management can be anticipated.

Reference is made to non-patented document titled as “Deep learning neural network with transfer learning for liver cancer classification” published in December 2022 by Nibras Mizour. This study describes an automated system using deep learning and transfer learning to detect liver tumors (HCC) in medical images. To improve accuracy with limited resources, they combined a pre-trained VGG-16 model for classification and a modified MobileNet-SSD model for tumor detection. The system achieved high accuracy (96%) in classifying cancerous areas, making it a promising tool for early HCC diagnosis and treatment decisions.

Another reference is made to non- patented document titled as “Cancer Cells Detection using OTSU Threshold Algorithm” published in December 2017 by Nalluri Sunny, Mithinti Srikanth, Kodali Eswar. Said research article uses an image processing techniques (OTSU thresholding & Watershed transformation) to detect lung cancer cells in CT scans, aiming for earlier diagnosis and improved treatment outcomes. However, this paper mainly focuses on simpler image processing techniques for lung cancer detection. Adding the VGG16, U-net architecture, and transfer learning based diagnostics is not utilized here for superior accuracy.

Existing inventions in this field lack in providing individual based strategy to increases the likelihood of a successful outcome while reducing unneeded treatments. Additionally, accurate staging helps in providing insightful prognosis, predicted course of the disease and possible outcomes which can better equipped the health care professionals to make decisions for treatment outcomes. The existing state of the art model's have limited capacity to precisely segregate tumors and also limited post-treatment monitoring. Also, due to the disease's complex biology, numerous clinical manifestations, and multiple natures, staging Hepatocellular Carcinoma (HCC) presents challenging issues. Accurate staging is more difficult because of the variability of HCC lesions in size and spatial distribution. Contrary to several cancers, HCC does not proceed linearly and frequently and involves multifocal liver lesions, making it difficult to use established staging techniques based on unifocal tumour models.

In order to obviate the drawbacks of the existing state of the art, there is a pressing need for a system that combines the strengths of deep learning and advanced preprocessing techniques for HCC staging. Said system should be capable of extracting both global and local features from medical images for comprehensive and better understanding of complex liver tumor traits. Furthermore, the system should be capable of benefiting from knowledge gained from a broader range of medical image datasets and surpassing conventional methods by offering a more robust and accurate solution for HCC staging.

OBJECT OF THE INVENTION
In order to overcome the shortcomings in the existing state of the art the object of the present invention is to provide a prognosis model for hepatocellular carcinoma staging.

Yet another objective of the invention is to provide a system capable of identification, segmentation and examination of tumor.

Yet another objective of the invention is to provide a system capable of extracting both global and local features from medical images for comprehensive and nuanced understanding of complex liver tumor traits.

Yet another objective of the present invention is to provide a preprocessing module capable to resize the MRI images to a standard resolution, apply Otsu’s binary thresholding to divide the images' foreground and background portions, and apply marker-based watershed segmentation to capture complicated tumor borders.

Yet another objective of the present invention is to provide a feature extraction module that is configured to extract features from the preprocessed MRI images.

Yet another objective of the present invention is to provide an ensemble module that uses modified design of VGG16 and U-Net deep learning architecture to achieve tumor segmentation which helps combining the strengths of both architectures to accurately identify and segment tumor regions in the liver MRI images.

Yet another objective of the invention is to provide a system capable of examining the changes that occur in the tumor regions, for accurate HCC staging and prediction.

SUMMARY OF THE INVENTION:
This invention presents a system for automated HCC staging and diagnosis that leverages an ensemble model (SEM). Said system incorporates the following modules:
 Data Module (SDM): This module is responsible for receiving magnetic resonance imaging (MRI) images of the liver.
 Preprocessing Module (SPM): The SPM performs several crucial tasks:
 Resizing: It resizes the received MRI images to a standard resolution for consistency.
 Segmentation: It utilizes Otsu's binary thresholding to segment the images, separating the foreground (potentially containing tumor) regions from the background.
 Refined Segmentation: Employing marker-based watershed segmentation, the SPM refines the segmentation of complex tumor borders, ensuring greater accuracy.
 Feature Extraction Module (SFEM): This module extracts informative features from the pre-processed MRI data. It comprises two sub-modules:
 Modified VGG16 Module (SFEM1): This sub-module extracts high-level global features from the data using a modified VGG16 architecture.
 Modified U-Net Segmentation Module (SFEM2): This sub-module leverages a modified U-Net architecture for local feature extraction and performs accurate tumor segmentation.
 transfer learning module is employed to leverage pre-trained weights from VGG16 and U-Net architectures to enhance said ensemble model's (SEM) performance.
 Ensemble Module (SEM): This module is the core of the invention. It integrates the modified VGG16 and U-Net architectures through a transfer learning approach. This combined ensemble model aims to achieve superior accuracy in HCC staging based on the extracted features.
This invention presents a novel system for automated HCC staging and diagnosis. It leverages an ensemble learning model that incorporates modified deep learning architectures. This approach offers the potential for improved accuracy, efficiency, and objectivity compared to existing technologies for HCC diagnosis.

BRIEF DESCRIPTION OF DRAWINGS
Figure 1 depicts a flow chart of the proposed methodology.
Figure 2 depicts Architecture of the ensemble model.
Figure 3 depicts MRI images and the corresponding results after Otsu’s Binary Thresholding.
Figure 4 depicts Segmentation result after Marker based watershed segmentation.
Figure 5 depicts (a) ROC curve of the ensemble model (b)Precision-Recall Curve of the Ensemble model
Figure 6 depicts lesion size distribution in various HCC Stages
Figure 7 depicts results after applying the ensemble model along with transfer learning.
Figure 8 (a) Training History of ensemble model (b) F1 score over various epochs.
Figure 9 (a) Confusion Matrix and (b) Correlation Matrix of the Ensemble model

DETAILED DESCRIPTION OF THE INVENTION WITH ILLUSTRATIONS AND EXAMPLES
While the invention has been disclosed with reference to certain embodiments, it will be understood by those skilled in the art that various changes may be made, and equivalents may be substituted without departing from the scope of the invention. In addition, many modifications may be made to adapt to a particular situation or material to the teachings of the invention without departing from its scope.

Throughout the specification and claims, the following terms take the meanings explicitly associated herein unless the context clearly dictates otherwise. The meaning of “a”, “an”, and “the” include plural references. Additionally, a reference to the singular includes a reference to the plural unless otherwise stated or inconsistent with the disclosure herein.

Table 1: Legend of Reference numerals
Ser no. Item reference Reference Numerals
1. System S
2. MRI Image Dataset SID
3. Preprocessing Module SPM
4. Otsu’s binary thresholding module SPM1
5. Marker based Watershed SPM2
6. Feature extraction module SEFM
7. Data module SDM
8. Ensemble module SEM
9. Evaluation module SEVM

Some of the technical terms used in the specification are elaborated as below:

VGG16: VGG16 is a well-known deep convolutional neural network with stacked convolutional and pooling layers as its basic architecture. It has demonstrated effectiveness in various image analysis applications while not being as in-depth as other recent models.

U-Net: U-Net is an architecture specifically designed to perform image segmentation tasks. Its expanding and contracting routes make it easier to precisely separate areas of interest in medical pictures.

Ensemble Model of VGG16 and U-Net: For improved HCC staging accuracy, the proposed ensemble model integrates VGG16 and U-Net, leveraging their complementary strengths in feature extraction and segmentation.

Figure 1 depicts flowchart of the present system’s methodology for hepatocellular carcinoma staging and diagnosis. Said system involves following steps:
 Data gathering: The present system utilized image dataset selected from but not limited to, MRI image dataset (SID) from patients diagnosed with hepatocellular carcinoma. This dataset (SID) offers overall liver health and HCC presentation. Said Magnetic resonance imaging (MRI) consists of a series of 3D images, known as slices, that show different views of the liver; each layer's image is 224 × 224 pixels. Most of the pictures in the collection are MRI scans, the imaging modality of choice for diagnosing hepatocellular carcinoma since they provide a high degree of contrast and detail for visualizing soft tissues, including tumors. The dataset includes different HCC stages, to denote the course of the disease.
 Preprocessing Module: The preprocessing sought to highlight necessary signals and reduce noise to improve the accuracy of HCC staging. Said images are first resized to a standard resolution to reduce computational complexity while preserving crucial features. The present invention utilizes following methods for preprocessing purposes:
 Otsu’s binary thresholding: Said method is used to effectively divide the images' foreground (tumour regions) and background portions. Said method improves later analyses by making tumour boundaries and other vital structures more visible. Said method helps improving the accuracy of subsequent segmentation tasks. This technique efficiently isolates parts of interest within the images by maximizing inter-class variance. Figure (3) depicts MRI images and the corresponding results after Otsu’s Binary Thresholding. Fig (3) represents the few original input images (rows 1,3,5) and the corresponding result (rows 2,4,6) after Otsu’s binary thresholding.
 Marker based Watershed segmentation: This method is particularly useful in capturing complicated tumor borders, improves segmentation procedure by segmenting clustered and related objects. In the context of medical image analysis, watershed segmentation can be particularly useful for delineating boundaries between adjacent structures or regions, such as separating tumors from surrounding healthy tissue in liver MRI scans. It can help create precise segmentation masks that outline regions of interest, making it a valuable tool for tasks like tumor detection and quantification. Figure (4) illustrates the input images (rows 1, 3, 5) and the corresponding results (2, 4, 6) after the marker-based watershed segmentation

Combining these strategies made it possible to precisely isolate tumor locations, improving the quality of the input data for the ensuing ensemble model.

Feature extraction module:
VGG16:
 VGG16 is a pre-trained convolutional neural network (CNN) designed for image classification tasks. In this pipeline, its primary role is feature extraction. VGG16 extracts hierarchical features such as edges, textures, and complex patterns from MRI input images after preprocessing. These extracted features are rich representations of the spatial and structural characteristics of the MRI dataset, which are crucial for identifying regions of interest. By leveraging transfer learning, VGG16 applies its learned weights to capture domain-specific features relevant to MRI segmentation and classification.

VGG16 features:
 Extracts global features such as textures, shapes, and patterns specific to the structures visible in MRI images.
 High-level abstract features like pathological markers that aid in identifying abnormalities.
 These features represent the broader spatial context but may not be localized.
U-Net:
 U-Net is an encoder-decoder architecture specifically designed for image segmentation tasks. In figure1, U-Net processes the preprocessed MRI images to perform segmentation, producing pixel-wise classifications to delineate regions of interest such as lesions or tumors. The segmentation output enhances the localization of specific structures in the MRI dataset, crucial for downstream tasks like classification and prediction. It complements VGG16 by focusing on spatial coherence and generating segmented masks.
U-Net Features:
 Focuses on pixel-level and localized features such as segmenting specific regions such as tumor boundaries or lesion areas.
 Capturing spatial relationships and fine-grained anatomical details.
 Ensures the features represent precise delineation for the ensemble to focus on critical areas.

 Ensemble module: The ensemble model integrates the abstracted (VGG16) and localized (U-Net) features to create a comprehensive feature set. This improves both the classification accuracy and the robustness of the predictions. To achieve the best performance in tumour segmentation, specific modifications are made to the design of VGG16 and U-Net, which is known for its ability to extract features, to optimize it for HCC staging. A key challenge was converting VGG16 for binary tumor segmentation (identifying tumor vs. non-tumor). To address this, the initial fully connected layers of VGG16 are modified. Said modification involved changing the output layer to generate a single-channel map showing the probability of a tumor being present at each pixel. This change allows the model's output to directly provide the information critical for HCC staging.
 U-Net's architecture with encoding and decoding paths is ideal for segmentation, However, it is needed to adapt its output for HCC staging. Instead of handling multiple classifications, U-Net architecture is modified to create a binary segmentation mask. This mask simplifies the task by indicating whether each pixel represents a tumor or healthy tissue, aligning perfectly with the study's goal of identifying cancerous areas. The majority of the customization effort focused on changing U-Net's output layer from multi-class segmentation to this binary tumor segmentation approach. This results in a mask where each pixel is labeled as either tumor or non-tumor, effectively highlighting the regions of interest for further analysis in HCC staging.
 Transfer learning module employed in the system (SEM3) is configured to leverage pre-trained weights from VGG16 and U-Net architectures to enhance the ensemble model's performance.

 Prediction and Evaluation Module:
The system depicted in Figure 1 provides outputs designed to assist healthcare professionals by presenting results in an interpretable and actionable manner, facilitating further diagnosis or treatment allocation. The U-Net segmentation generates pixel-wise labeled maps of MRI images, highlighting regions of interest such as tumors, lesions, or abnormal tissues. This visual output enables clinicians to accurately localize abnormalities, assess the size, shape, and location of affected regions, and plan interventions like surgery or radiation therapy. Additionally, the classification network, leveraging features extracted by VGG16 and fully connected layers, produces probabilistic labels such as benign versus malignant or tumor stage classification, offering quantitative diagnostic insights. These outputs guide further diagnostic steps, such as ordering additional imaging or biopsies. By integrating the outputs from U-Net and VGG16, the ensemble model provides a comprehensive prediction that combines localiz
The proposed system enhances interpretability through visual outputs such as MRI images overlaid with segmented regions and annotated with classification results. Heatmaps or attention maps may also indicate areas of high diagnostic relevance, while detailed, clinician-friendly reports summarize key findings such as the size, shape, and volume of abnormal regions, along with probabilistic diagnoses and confidence scores. These reports may also suggest next steps for treatment or further investigation. Quantitative metrics, such as tumor volume or growth rate, and predictive analytics, like survival rates or treatment response, further facilitate treatment allocation, enabling personalized care. Results are presented through an intuitive, interactive interface, allowing healthcare professionals to explore images, adjust segmentation thresholds, or compare findings over time, with seamless integration into existing healthcare systems such as PACS. Overall, the system improves the accuracy of identifying and characterizing abnormalities, saves time by automating repetitive tasks, reduces subjectivity in diagnosis, and enhances treatment planning to improve patient outcomes.

The following detailed description provides insights into each component of the present invention:

The dataset used in this invention is MRI images of the liver. These images were meticulously labeled with detailed information about hepatocellular carcinoma (HCC). To ensure the images accurately reflect real-world situations, the dataset captured variations in tumor size, location, and imaging conditions. Furthermore, data augmentation techniques were implemented to increase the dataset's diversity and improve the model's ability to generalize across different scenarios. These techniques involved random rotations, horizontal flips, and scaling, simulating various viewing angles and patient positions during imaging. Data augmentation effectively boosted the dataset size and enhanced the model's robustness to such variations. Following this, proper data preprocessing steps were undertaken. All images were resized to a uniform 256x256 pixel resolution for consistency. Additionally, intensity normalization was performed to standardize the pixel values across different scanner settings, further improving the model's ability to learn effectively from the data. Finally, meticulous data curation ensured the integrity and completeness of the dataset by eliminating any missing or damaged images.

Said dataset is split into training and validation sets, train our ensemble model, predict probabilities, and then calculate and plot the ROC curves for both the training and validation sets, illustrated in figure 5 (a).

Accurate model training and evaluation depend on the dataset preparation stage. The dataset's quality, consistency, and applicability for the challenging hepatocellular carcinoma (HCC) staging task were rigorously examined. Beginning with a rigorous data cleaning procedure, the preprocessing process set out to remove any noise or artifacts that would impair proper analysis. After that, pixel intensity values were standardized to eliminate inconsistencies from different imaging situations.

For said purpose, a key component of our preprocessing technique, Otsu's binary thresholding module (SPM1) plays a crucial role in boosting data quality. Said module (SPM1) automatically divides pixels into foreground (tumour) and background regions by an ideal threshold. Said module (SPM1) accurately isolates the tumour boundaries by accurately discriminating between these locations. This binary thresholding strengthens the basis for subsequent analysis phases by minimizing the impact of background noise while enhancing tumour delineation.

The steps involved in Otsu's binary thresholding Module (SPM1) are as follows:
1. Histogram Calculation: Let H(i) be the image's histogram, representing the frequency of occurrence of each intensity value “i” in the picture.
2. Normalization: Normalize the histogram by dividing each bin by the total number of pixels in the image so that the values in the histogram sum up to 1, where N is the total number of pixels in the image.
p(i)=H(i)/N (1)
3. Cumulative Sum: Compute the cumulative sum of the normalized histogram from the minimum intensity value to the maximum intensity value:
p(i)=∑p(k), for=k=0 to i (2)
4. Cumulative Mean: Calculate the cumulative mean up to each intensity value
μ(i)=∑k*p(k) for k=0 to i (3)
5. Global Mean: Compute the global mean of the image, which is the weighted sum of the cumulative mean values, with the weights being the cumulative histogram values:
μ_t=∑i*p(i) for i=0 to L-1 (4)
where L is the number of intensity levels.
6. Between-class Variance: Compute the between-class variance for each intensity value as
σ_B^2(i)=[μ_t*P(i)-μ(i)]^2/[P(i)*(1-P(i))]for i=0 to L-1 (5)
7. Optimal Threshold: Find the intensity value that maximizes the between-class variance, which is the optimal threshold value:
T_opt=argmax(σ_B^2(i)), for i=0 (6)
8. Thresholding: Apply the threshold value to the image to separate the pixels into two groups, foreground and background
9. If I: (x,y)=T_opt,then I(x,y)=1(foreground pixel) (8)
where I (x, y) represent the pixel value at location (x, y) in the image.

Application of marker-based watershed segmentation module (SPM2) complements Otsu's binary thresholding module (SPM1). By precisely segmenting nearby structures, this method sharpens the tumour boundaries. The segmentation process is guided by markers, or "seed points," making it easier to distinguish between tumour patches and other anatomical features. The end result is a more accurate tumour segmentation with the spatial context required for precise HCC staging.
Said marker-based watershed segmentation Module (SPM2) performs following steps:
 Marker Generation: Choosing markers for the image. These markers could be obtained by different methods such as user input, thresholding, or other image processing techniques. Let M (x, y) be the marker image, which is a binary image where the foreground pixels represent markers and the background pixels represent the rest of the image.
 Distance Transform: Compute the distance transform of the marker image. The distance transform assigns each pixel in the picture a value that represents the distance to the nearest marker. Let D (x, y) be the distance transform of the marker image.
 Watershed Segmentation: Apply the watershed transformation to the distance transform image. The watershed transformation is a flooding algorithm that starts at the markers and floods the image until the flooding fronts meet. The boundaries between the flooded regions define the watershed lines. Let W (x, y) be the watershed image, a binary image where the foreground pixels represent the watershed lines and the background pixels represent the flooded regions.
 Result Generation: Combine the marker image and the watershed image to obtain the segmented image. The segmented image is obtained by assigning each pixel to the nearest marker. Let S (x, y) be the segmented image.

The mathematical equations for the above steps are as follows:
Marker Generation:M(x,y)=1 if the pixel at (x,y)is a marker.0 otherwise (9)
Distance Transform:D(x,y)=min⁡(dist(M(x,y),M(i,j))),for all(i,j) (10)
in the image, where dist. is the Euclidean distance between two points.

Watershed Segmentation: Initialize a priority queue Q with all the marker pixels in the distance transform image. While Q is not empty, do the following:
• Pop the pixel (x, y) with the smallest distance from Q.
• For each neighbor (i, j) of (x, y), do the following:
Case1: If (i, j) is not already in the queue, add it to the queue and assign it the same label as (x, y).
Case 2: If (i, j) is not labeled and has a lower distance than any other unlabeled pixel, assign it the label of (x, y) and add it to the queue.
Case 3: If (i, j) is not labelled and has a higher or equal distance to any other unlabelled pixel, mark it as a watershed pixel.

The result of the watershed segmentation is the watershed image W (x, y), which is a binary image.
Result Generation:S(x,y)=argmin(D(x,y,M(i,j))), for all (i,j) that are markers…(11)

The segmented image S(x,y) is the final result of the marker-based watershed segmentation.

In figure 6 the histograms display the distribution of lesion size values within each HCC stage. This enables us to see the skewness, dispersion, and core tendencies of each stage's feature. Sharp variations in the histograms may imply that lesion size is a useful characteristic for identifying different stages. As HCC lesions move from benign or early stages to more advanced stages, including tertiary stages, they frequently enlarge. This expansion is frequently linked to the unchecked growth of cancer cells.

VGG16 and U-Net, two deep learning architectures are modified to improve their performance in tumor segmentation for precise HCC staging.
 VGG16 Modification: VGG16 a deep learning architecture, known for its feature extraction capabilities, required adjustments for HCC staging. The challenge was to adapt it for binary tumor segmentation (identifying tumor vs. non-tumor). This was achieved by reconfiguring the initial fully connected layers and changing the output layer. The new output layer produces a single-channel probability map. Each pixel's value in this map represents the probability of a tumor being present. This modification allows the model's output to directly provide the information critical for HCC staging, which is the presence or absence of tumors.

 U-Net Modification: While U-Net's architecture with encoding and decoding paths was ideal for segmentation, its output layer needed adaptation for HCC staging. Originally designed for multi-class segmentation, U-Net is modified to create a binary segmentation mask. Said mask simplifies the task by indicating whether each pixel represents a tumor or healthy tissue. The customization is done by changing U-Net's output layer from multi-class segmentation to this binary tumor segmentation approach. Each pixel in the resulting mask is labeled as either tumor or non-tumor, effectively capturing the regions of interest for further analysis in HCC staging.

 Ensemble Model: By incorporating these modifications to VGG16 and U-Net, the ensemble model is well-suited for HCC staging. The reconfigured output layer of VGG16 and the binary segmentation mask of U-Net ensure the model's predictions directly align with the goal of identifying and demarcating tumor regions. These adjustments demonstrate the focus on achieving high accuracy and allow the ensemble model to segment liver tumors more precisely. Further details about the network architecture of the proposed ensemble model can be found in Table 1.

Table 1 depicts Network structure of the ensemble model
Layer Name Input Shape Output Shape Activation Function
Input (VGG16) (224,224,3) (224,224,3) None
VGG16 Block 1 (224,224,3) (56, 56, 128) ReLU
VGG16 Block 1 (56, 56, 128) (28, 28, 256) ReLU
VGG16 Block 3 (28, 28, 256) (14, 14, 512) ReLU
Input (U-Net) (256, 256, 3) (256, 256, 3) ReLU
U-Net Encoding 1 (256, 256, 3) (128, 128, 64) ReLU
U-Net Encoding 2 (128, 128, 64) (64, 64, 128) ReLU
U-Net Encoding 3 (64, 64, 128) (32, 32, 256) ReLU
Concatenation Layer (14, 14, 512 + 256) (14, 14, 768) ReLU
Convolution Layer (14, 14, 768) (14, 14, 256) ReLU
Convolution Layer (14, 14, 256) (14, 14, 128) ReLU
Convolution Layer (14, 14, 128) (14, 14, 1) Sigmoid

Figure 7 shows the results after applying the ensemble model along with transfer learning. The image consists of six columns. The input images and predicted output results are represented in columns (1,3,5) and (2,4,6) respectively.

To plot the Receiver Operating Characteristic (ROC) curve for both the training set and validation set, first, we calculate the actual positive rate (TPR) and false positive rate (FPR) for different thresholds of our classifier's predicted probabilities. The dataset is split into training and validation sets, train our ensemble model, predict probabilities, and then calculate and plot the ROC curves for both the training and validation sets, illustrated in Figure 5 (a). We must calculate precision and recall values for different probability thresholds of our classifier's predicted probabilities to draw the Precision-Recall curve. The dataset is split into training and validation sets, trains our ensemble model classifier, predicts chances, and then calculates and plots the Precision-Recall curves for both the training and validation sets in Fig. 5 (b).

To assess the ability of the present system to predict and distinguish between tumor and non-tumor regions, we evaluated its Area Under the Curve (AUC) value. The achieved AUC of 0.94 demonstrates the model's exceptional aptitude for discrimination. A high AUC value signifies the model's strong ability to differentiate between tumor and non-tumor regions. This distinction is crucial in hepatocellular carcinoma, where accurate diagnosis is essential. Because of this impressive AUC, the ensemble model excels at both detecting malignancies and precisely delineating their boundaries. This level of precision empowers clinicians to develop individualized treatment plans, evaluate prognoses, and optimize patient care processes.

The proposed ensemble model's loss value is presented beside the AUC value. An indicator of model convergence and accuracy, the loss value, was calculated to be 0.08. This low loss number highlights the model's capacity to faithfully simulate the tumour locations in the real-world during training. Figure 8 (a) provides a visualization of the training history of our ensemble model across different epochs. It shows how loss and accuracy change as the model undergoes training over multiple generations. The green line with markers ("o") and the red line with tags ("o") at each epoch represents the model's training and validation accuracy, respectively. The degree to which the model's predictions coincide with the actual target values during training is measured by training loss. The model's performance on omitted validation data is calculated by validation loss. As the model improves at making predictions, training accuracy—the percentage of instances on the training data that were predicted correctly—should rise over time. In Fig. 8 (b), each epoch's F1 score is plotted against the number of epochs. This can help us visualize how the F1 score changes over the course of training iterations and assess your model's performance in terms of precision and recall.

In the context of a classification issue, a confusion matrix is a table that lists how well a classification algorithm performed. It displays the algorithm's predictions across various classes compared to the correct labels. Understanding how well your model performs for each category and spotting misclassification trends are also possible uses for the confusion matrix. The actual classes are represented by the rows in the confusion matrix, while the columns represent the anticipated classes. The number of accurate predictions for each category is shown by the diagonal elements, which go from top left to bottom right. Off-diagonal elements represent misclassifications. For instance, 5 class "benign" instances were incorrectly labeled as "primary."

The sum of each row represents the total examples of that class, while the sum of each column represents the actual instances that were predicted to belong to that class. Figure 9 (a) illustrates the confusion matrix of the ensemble model. The graph sheds light on the generalization and learning processes in your model. In a perfect world, training loss would fall, validation loss would decrease (up to a point), and training and validation accuracy would rise or stabilize. Indicators of overfitting and poor generalization to new data include increased validation loss. This graph is a standard tool for tracking training progress and aids in deciding when to halt training or whether the model is overfitting. Figure 9 (b) is the ensemble model heatmap that visualizes the HCC dataset's correlation matrix.
, C , Claims:WE CLAIM:
1. A system for hepatocellular carcinoma staging and diagnosis comprising:
 data module (SDM) configured to receive MRI image dataset (SID) of the liver;
 preprocessing module (SPM) configured to remove any noise or artifacts that would impair proper analysis of said received MRI images, said preprocessing module (SPM) comprising:
 Otsu’s binary thresholding module (SPM1);
 Marker based watershed segmentation (SPM2);
 feature extraction module (SFEM) configured to extracting information from said preprocessed MRI data, said feature extraction module (SFEM) comprising:
 VGG16 based feature extraction module (SFEM1);
 U-net segmentation module (SEFM2);
 ensemble module (SEM) to predict the stage of a tumor based on the extracted features, said ensemble module (SEM) comprising:
 modified VGG16 module (SEM1);
 modified U-net module (SEM2);
 transfer learning module (SEM3);
characterized in that,
wherein said system (S)
 said preprocessing module (SPM) configured to resize said MRI images to a standard resolution.
 said Otsu’s binary thresholding module (SPM1) configured to divide the images' foreground and background portions.
 said marker-based watershed segmentation is employed to capture complicated tumor borders.
 data augmentation techniques are used to increase the dataset's diversity and improve the model's ability to generalize across different scenarios.
 said ensemble module (SEM) developed by incorporating modifications to VGG16 and U-Net, to create system (S) capable of predicting and identifying tumor, and its progression.
 transfer learning module (SEM3) configured to leverage pre-trained weights from VGG16 and U-Net architectures to enhance the ensemble model's performance.
 transfer learning, modified U-Net, and VGG16 architecture are strategically combined to create said ensemble model that exceeds individual architectures' limitations.

Said system (S) capable of capturing the complex nuances of liver tumor imaging, which ensures that present system can precisely collect the range of data needed for exact HCC staging prediction.

2. The system as claimed in claim 1, wherein said image dataset (SID) are selected from but not limited to Magnetic Resonance Imaging (MRI) image of the liver.
3. The system as claimed in claim 1, wherein said preprocessing Module (SPM) sought to highlight features and size of the tumor and reduce noise to improve the accuracy of HCC staging.
4. The system as claimed in claim 2, wherein said preprocessing module (SPM) performs operations such as resizing, normalizing, and undergo segmentation on said MRI dataset (SID) using Otsu's thresholding (SPM1) and marker-based watershed segmentation module (SPM2) to isolate tumor regions within said liver MRI.
5. The system as claimed in claim 2, wherein said Otsu’s binary thresholding module (SPM1) improves the accuracy by isolating parts of interest within the images by maximizing inter-class variance.
6. The method of claim 7, wherein said marker-based watershed segmentation module (SPM2) is configured to improve the segmentation procedure by segmenting clustered and related objects.
7. The system as claimed in claim 1, wherein said data augmentation techniques includes but are not limited to random rotations, horizontal flips, and scaling, simulating various viewing angles and patient positions during imaging.
8. The system as claimed in claim 1, wherein said ensemble module (SEM) is configured to use modified design of said VGG16 and U-Net deep learning architecture to achieve tumor segmentation.
9. The system as claimed in claim 8, wherein said wherein said VGG16 module is modified (SEM1) for binary tumor segmentation, the modifications comprising:
 modifying the initial fully connected layers of the VGG16 network; and
 replacing the original output layer of the VGG16 network with a single-channel output layer configured to generate a probability map, wherein each pixel value in the probability map represents the probability of a tumor being present at the corresponding pixel in the preprocessed MRI data.
10. The system as claimed in claim 8, wherein said U-Net network is modified for binary tumor segmentation, the modifications comprising:
 modifying the final layers of the U-Net network's decoder path; and
 replacing the original output layer of the U-Net network with a single-channel output layer configured to generate a binary segmentation mask, wherein each pixel value in the mask is labeled as either 1 (tumor) or 0 (non-tumor).
11. A method for hepatocellular carcinoma (HCC) staging and diagnosis using a computer system, the method comprising:
 receiving a magnetic resonance imaging (MRI) image dataset of the liver;
 pre-processing the received MRI images, the pre-processing comprising:
 resizing the preprocessed images to a standard resolution;
 segmenting the images by separating foreground (potentially containing tumour) and background regions using Otsu's binary thresholding;
 segmenting complex tumour borders within the foreground regions using marker-based watershed segmentation;
 extracting informative features from the preprocessed MRI data using a feature extraction module comprising:
 VGG16-based module configured to extract high-level global features;
 U-Net segmentation module configured to extract local features and perform accurate tumour segmentation;
 predicting the stage of a tumour based on the extracted features using an ensemble model comprising:
 a modified VGG16 module adapted from the VGG16 architecture to improve its effectiveness for HCC staging;
 a modified U-Net module adapted from the U-Net architecture to enhance its ability for accurate tumor segmentation in the context of HCC staging;
 a transfer learning module configured to leverage pre-trained weights from VGG16 and U-Net architectures to enhance the ensemble model's performance in HCC staging;