A Robust and Lightweight Deep CNN for Liver Disease Classification in CT and MRI Using Enhanced Full-Image Processing

Liver cancer is one of the leading causes of cancer-related deaths globally. Despite advancements in treatment, the disease still exhibits high mortality rates, largely due to challenges in early and accurate diagnosis. Medical imaging, particularly CT and MRI, remains the cornerstone of liver disease evaluation. However, conventional diagnosis methods are often time-consuming and subject to inter-observer variability, necessitating automated and intelligent systems based on deep learning [1-3]. CT and MRI imaging techniques are the primary tools for identifying hepatic lesions through detection, diagnosis, and monitoring. The challenge of underdiagnosis emerges because of subtle tissue contrast differences and noise from variations in anatomical structures, which interfere with proper treatment. Using traditional methods with prior knowledge and tissue properties usually results in unpredictable results, but deep learning models built on ResNet-50, AlexNet, and GoogLeNet achieve better segmentation accuracy. Despite the advancements in medical imaging, these improvements require higher computational resources [4, 5].

Abdominal CT scans are a valuable tool for the early detection and characterization of various liver cancers. They provide detailed information on the size, shape, and location of tumors within the liver and surrounding abdomen, along with the vascular anatomy. Accurate lesion characterization is crucial for treatment planning. However, radiologists face the time-consuming task of manually detecting and segmenting lesions within complex 3D CT images containing numerous lesions. This highlights the need for computer-aided analysis to assist physicians in liver metastasis detection and size evaluation in CT scans. Automatic segmentation of liver lesions and parenchyma remains challenging due to lesions' varying contrast enhancement patterns and the inherent variability in perfusion and scan timing, leading to poor image contrast between these tissues [6, 7].

This study explores the development of a deep-learning model for liver disease classification through a comparative analysis of various classifiers. A confusion matrix was utilized to assess each model's classification performance using multiple evaluation metrics. Given the complex clinical features of liver cancer histopathological images and the heterogeneous training data, a CNN model with superior architecture is proposed for tumor detection in liver CT and MRI images to address the challenges of feature extraction. The key contribution of our work lies in designing a scalable CNN model that can effectively handle both large, heterogeneous datasets and smaller datasets.

Despite notable advancements in liver image classification using CNNs, many existing models still depend on high computational resources, manual region selection, or are trained on limited, modality-specific datasets. Moreover, few frameworks simultaneously support CT and MRI images in a unified architecture. To address these limitations, this study proposes a robust and lightweight CNN model that processes entire liver images, eliminating the need for patch-based input and reducing redundant computations. The architecture integrates optimized preprocessing techniques, including HU filtering and CLAHE, to enhance tissue contrast and suppress irrelevant regions. Designed for small and large heterogeneous datasets, the model achieves classification accuracy exceeding 99%, combining high diagnostic performance with computational efficiency. This balance of adaptability, precision, and speed represents the core contribution of this research.

The paper is structured as follows: Section 2 reviews related work; Section 3 outlines the methodology; Section 4 presents the results and comparative analysis of the proposed method; and Section 5 provides the conclusion.

This section outlines the proposed CNN-based framework for the classification of liver diseases from CT and MRI images. The overall methodology is illustrated in Figure 1 and consists of three main stages: preprocessing, feature extraction, and deep learning-based classification. The prescribed workflow begins with the image preprocessing phase, where unhindered compatibility with the CNN is ensured by transforming the DICOM-format CT images into the JPG format. Furthermore, the HU intensity scale is also limited to values in the range of -75 to 150 to emphasize the tissue densities of the liver and exclude irrelevant regions such as bones and air in CT images. After windowing the HU images, CLAHE is applied to both the HU images and the JPG version of the MRI images to enhance contrast and outline the tumor edges alongside the physiological tissue appearances. Other preprocessing steps comprise intensity thresholding to leave intensities that are clinically significant, image resizing to a standard spatial resolution of 100 pixels by 100 pixels, and normalization of pixel values across all datasets to a standard pixel value. All the images are reconstructed into three channels before they are analyzed to meet the input requirements of the convolutional neural network.

1.png

Figure 1. Block diagram for the proposed model

Data augmentation is applied to expand the dataset and improve generalizability through rotation, flipping, and brightness adjustments. The refined data is split into 65% for training, 10% for validation, and 25% for testing. The CNN model is trained to extract hierarchical features from the liver images, automatically learning texture, shape, and structural patterns associated with various liver conditions. Model parameters are optimized iteratively to minimize classification error, based on validation metrics that include accuracy, sensitivity, and specificity.

Upon completion of training, the model is deployed to classify liver images into three categories: cancer, tumor, and normal. The learned feature representations guide the classification, enabling robust performance across heterogeneous image modalities. The proposed pipeline integrates advanced preprocessing, high-quality imaging datasets (TCGA-LIHC, 3D-IRCADb-01, and LiTS17), and an efficient CNN architecture, offering a reliable approach for automated liver disease detection and diagnosis.

3.1 The liver datasets preparation

This study utilized three publicly available liver imaging datasets: 3D-IRCADb-01, LiTS17, and TCGA-LIHC [18-20]. These datasets contain abdominal CT scans originally stored in the DICOM (.dcm) format. However, DICOM files are not natively supported by most deep learning libraries. All DICOM images were converted to JPEG (.jpg) format to address this issue. The conversion process was implemented using MATLAB. A custom conversion pipeline was applied and executed in two stages. The main script iterated through a list of DICOM file identifiers. A helper function (dicom2image.m) was called for each file. This function uses MATLAB’s dicomread to load the image data. The image was normalized using mat2 gray, scaled to an 8-bit representation using uint8, and saved in JPEG format using imwrite. This procedure ensured that the images were preserved with high visual fidelity while being formatted for compatibility with deep learning models. The converted images were resized and stored in a consistent format and resolution, enabling streamlined training and testing of the custom-designed convolutional neural network.

The 3D-IRCADb-01 dataset provides CT scans from 20 patients (10 male, 10 female), ranging from 260 to 2800 axial slices per scan at a resolution of 512×512. The LiTS17 dataset comprises 130 training and 70 testing volumes for the liver tumor segmentation benchmark.

Our decision to rely heavily on well-documented public data, including TCGA-LIHC, 3D-IRCADb-01, and LiTS17, to conduct this fundamental study was driven by interconnected issues. Top among these is the provision of standardized procedures for acquisition, which, in addition to making the direct comparison of the results to current state-of-the-art procedures a less daunting task, also makes the reproducibility of our findings that much easier. Nevertheless, we are mindful of the shortcomings that real-life clinical circumstances impose with heterogeneous scanner manufacturers, varying protocols, different demographics of patients, and variable clinical presentations of diseases, which can lead to misinterpretations.

3.2 Preprocessing

All CT images underwent a systematic pre-processing pipeline to enhance structural clarity and improve the reliability of feature extraction for downstream classification tasks. The initial step involved converting raw DICOM pixel values to HU, which represent a standardized radiodensity scale used in computed tomography. This conversion is crucial for ensuring consistency across imaging devices and clinical settings. The transformation is performed using a linear mapping extracted from DICOM metadata:

$\begin{gathered}H U=RawValue \times RescaleSlope+RescaleIntercept\end{gathered}$               (1)

This enables meaningful interpretation of tissue densities; the range of -75 to 150 HU was deliberately chosen based on established clinical practices for liver CT imaging [21, 22]. After conversion, intensity windowing was applied to isolate and emphasize relevant anatomical structures. Voxels outside the selected liver window were clipped, and the resulting image was rescaled to a normalized range of [0,1], preparing it for further enhancement.

To amplify local contrast and expose subtle tissue boundaries, CLAHE was applied to the windowed HU images. CLAHE divides the image into contextual regions (tiles), performs histogram equalization within each, and then combines the results using bilinear interpolation. To prevent over-enhancement of noise, a clip limit (0.01) was set, controlling the maximum slope of the cumulative distribution function within each tile. The intensity histogram h_m,n(r) for a tile T_m,n was computed as:

$h_{m, n}(r)=\sum_{(x, y) \in T_{m, n}} \delta(I(x, y)-r)$        (2)

where, δ is the Kronecker delta function, excess pixels above the clip limit were redistributed to maintain contrast without introducing artifacts.

The 8×8 tile size ensured that enhancement was localized to small, clinically relevant regions, while the modest clip limit of 0.01 was effective in enhancing the visibility of subtle tumor boundaries and textural details without amplifying background noise. This combination proved essential for normalizing the diverse datasets and preparing them for the subsequent feature extraction stages. It is important to note that using other parameter values for CLAHE, such as a clip limit of 0.02 with an 8×8 tile size (resulting in 95.6% accuracy), or a clip limit of 0.01 with a 4×4 tile size (90.1% accuracy) or 16×16 tile size (85.7% accuracy), consistently led to sub-optimal performance compared to our chosen parameters.

Following contrast enhancement, a 2D median filter was employed to suppress high-frequency noise and improve overall image quality. This non-linear filter replaces each pixel with the median of its neighboring values within a fixed window (3×3), making it highly effective against impulsive noise while preserving anatomical boundaries. Compared to linear smoothing filters, the median filter maintains edge sharpness and structural fidelity, which are essential in medical imaging analysis.

The integrated application of HU normalization, intensity windowing, CLAHE, and median filtering yielded images with improved clarity, balanced contrast, and reduced noise artifacts. This pre-processing sequence proved particularly beneficial for subsequent classification tasks, as it enhanced the visibility of diagnostically significant features while suppressing irrelevant variability. Figure 2 shows the liver image before and after the pre-processing using HU, CLAHE, and a median filter.

2.png

Figure 2. Preprocessing state of the CT image before and after applying HU, CLAHE, and the median filter

3.3 Proposed CNN design

The suggested CNN model is designed to extract high-level features from 100×100-pixel grayscale medical images. A proposed CNN architecture for liver tumor classification comprised six convolutional blocks and a fully connected layer. Each convolutional block consisted of a convolutional layer with more learnable filters (16, 32, and 64 for the subsequent four layers), capturing more abstract features from the images. The architecture's deeper convolutional layers are designed to capture high-level discriminative features, such as morphological or textural patterns that distinguish pathological abnormalities from normal regions. These layers are sequentially followed by batch normalization, rectified linear unit (ReLU) activation, and max-pooling operations to optimize feature robustness and computational efficiency. The network architecture integrated four max-pooling layers, each contributing to the progressive downsampling of feature maps. Full and convolutional layers followed a ReLU activation function and batch normalization to enhance training stability and make the network more effective. By combining them, it brings in non-linearity, allowing for the network to learn complex patterns. To mitigate overfitting, dropout regularization with a probability of 0.1 was applied to the fully connected layer. Finally, softmax activation was utilized to compute low-risk probabilities. The model was optimized using the Adam (Adaptive Moment Estimation) optimizer with a learning rate of 0.001 and trained over 25 epochs, allowing fast convergence and high performance across diverse datasets. The proposed model's structure is illustrated in Figure 3.

3.png

Figure 3. The proposed CNN structure

3.4 Feature extraction

Mathematical transformations play a crucial role in extracting hidden or subtle patterns from medical images that are not easily discernible in their raw form. Among the most prominent techniques, the Discrete Wavelet Transform (DWT) has demonstrated high efficacy in image analysis, particularly in medical image feature extraction. In the context of this study, 2D-DWT was employed on liver CT images to extract meaningful texture and structural features. The wavelet transform is capable of decomposing an image into a set of sub-bands that represent different frequency components and orientations, capturing both spatial and spectral characteristics. The 2D-DWT is applied in a separable manner: first across the rows and then across the columns of the image, effectively generating four sub-band components — approximation (LL), vertical details (LH), horizontal details (HL), and diagonal details (HH).

The decomposition is mathematically defined using low-pass and high-pass filters, denoted by h[n] and g[n], respectively. For a 1D discrete signal, the approximation and detail coefficients at the first level of decomposition are obtained as:

$\begin{gathered}\text { Low-pass (approximation):} a[k]=\sum_n x[n]. h[2 k-n]\end{gathered}$             (3)

High-pass (detail): $d[k]=\sum_n x[n] \cdot g[2 k-n]$             (4)

In 2D-DWT, this decomposition is extended to two dimensions by applying the 1D-DWT first along the image rows and then along the columns, resulting in four sub-band images which can be mathematically represented by:

$L L(i, i)=\sum_m \sum_n I(m, n) \cdot h(2 i-m) \cdot h(2 j-n)$               (5)

$L H(i, j)=\sum_m \sum_n I(m, n) \cdot h(2 i-m) \cdot h(2 j-n)$              (6)

$H L(i, j)=\sum_m \sum_n I(m, n) \cdot h(2 i-m) \cdot h(2 j-n)$                (7)

$H H(i, j)=\sum_m \sum_n I(m, n) \cdot h(2 i-m) \cdot h(2 j-n)$                   (8)

where, I(m,n) is the input image.

In this study, we utilized the Haar wavelet, a commonly used wavelet function defined by its simple step-like structure, suitable for detecting sharp transitions and edges in CT images. We performed both one-level and two-level decompositions, in which only the LL sub-band is recursively decomposed to achieve a multiscale representation. The 2D-DWT-based feature extraction provided a multi-resolution and compact description of liver CT images, enabling the identification of both global structures and fine textures relevant to liver pathology.

3.5 Training

The proposed CNN model learns image features through multiple convolutional and pooling layers, followed by a fully connected layer for classification. To optimize performance, hyperparameters were tuned via grid search over discrete values, as detailed in Table 1, with accuracy used as the primary selection criterion. The model was trained using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and for 25 epochs. A dropout rate of 0.1 was applied to reduce overfitting. During training, weights were updated iteratively through forward and backward propagation to minimize loss and ensure stable convergence.

Table 1. Hyperparameter tuning and optimization values

A data split of 65% for training, 10% for validation, and 25% for testing was adopted to provide a sufficient training volume while maintaining a robust and unbiased test set for final evaluation.

Our approach involved a series of practical experiments where different combinations of hyperparameters were tested to ensure a stable training process, avoiding issues such as overfitting identify the configuration that yielded the optimal balance between model performance (primarily accuracy) and generalization capability (minimizing the gap between training and validation curves).

We systematically varied critical hyperparameters, including the learning rate, batch size, and optimizer settings. The impact of these variations was observed through the model's convergence behavior, final accuracy, and the learning curves (training vs. validation loss/accuracy over epochs). Configurations that led to rapid convergence without significant overfitting, and ultimately achieved the highest validation accuracy, were selected.

Data augmentation played a critical role in enhancing the model's generalizability and robustness. By introducing transformations such as rotation, flipping, and shearing, the model was exposed to varied orientations and intensities, reducing overfitting and improving classification accuracy across all datasets. Although the class distributions in the benchmark datasets used were relatively balanced, augmentation also helped mitigate minor class imbalances, especially for rare tumor subtypes. Regarding the architecture illustrated in Figure 4, each convolutional layer employed 3×3 filters with a stride of 1 and 'same' padding to preserve spatial dimensions. Max-pooling layers with 2×2 filters and a stride of 2 were applied after specific convolutional blocks to downsample the feature maps progressively. The model operates on 2D image slices; therefore, standard 2D convolutions were used instead of 3D convolutions. While volumetric CT/MRI data were available, slices were extracted individually to reduce complexity and maintain training efficiency without compromising diagnostic relevance.

4.jpg

Figure 4. The training progress of a deep learning model designed

Table 5. Performance comparison between the selected models

A fully connected convolutional neural network (FC-CNN) was introduced in research [24] for the segmentation and classification of hepatic tumors in CT scans, achieving 99.11% accuracy on the 3D-IRCADb-01 dataset with noted computational efficiency. However, the absence of sensitivity and specificity metrics raises concerns regarding its diagnostic reliability. Moreover, the model was evaluated on a single dataset, limiting its generalizability across diverse imaging protocols and patient populations.

The authors in their research study [25] developed a spliced deep learning framework that consists of CNNs and EfficientNetV2B3 for IDC vs. non-IDC classification in WSIs. The model's good performance was attributed to its use of transfer learning and multi-stage preprocessing pipeline. Our best model performed similarly well with 96.3% accuracy, 93.4% precision, 86.4% recall, and an F1-score of 89.7%.

The authors in the study [26] introduced a hybrid model based on a Differential CNN and a KELM classifier for the detection of HB tumor. Evaluated on LiTS17 and 3D-IRCADb-01 datasets, it attained 98.72% accuracy, 98.24% recall, and 98.52% sensitivity. However, the model is computationally expensive with KELM and has a bit inferior specificity to state-of-the-art CNNs, which may cause misdiagnosis of benign anomaly. 

The authors in the study [27] proposed the UNet70 model, an enhanced version of the UNet architecture featuring 70 convolutional layers and hybrid residual connections. Designed to address limitations in tumor heterogeneity and small lesion detection, the model processed contrast-enhanced CT and MRI scans. UNet70 achieved 94.58% accuracy, a Dice score of 94.73%, and 97.50% sensitivity on a multi-institutional dataset.

A CNN-based framework was proposed in the study [28] for hepatic malignancy detection, leveraging transfer learning and feature fusion from multiple pretrained models, including ResNet-50 and DenseNet-121. Evaluated on LiTS17 and 3D-IRCADb-01 datasets, the model achieved 99.5% accuracy, with a precision of 0.864 and a recall of 0.979.