Image processing has emerged as one of the most rapidly evolving scientific fields in recent decades, with its applications expanding continuously across a wide range of industries. Today, there are very few industrial sectors that do not benefit from image processing techniques. Among these, the medical industry has been one of the most significant beneficiaries of advances in image analysis and computer-aided diagnosis systems. Medical imaging devices represent an important branch of healthcare technology, integrating sophisticated hardware and software components to assist clinicians in disease diagnosis and treatment planning. As technological advancements continue to reduce dependence on manual human intervention, medical imaging systems are becoming increasingly intelligent and efficient [1].

Computed Tomography (CT) imaging has become an indispensable diagnostic tool for the detection and monitoring of abnormalities in various organs, including the heart, liver, kidneys, intestines, and vascular structures. The relatively low operational cost, absence of proven adverse effects for both patients and clinicians, high imaging speed, and the capability to capture moving anatomical structures such as cardiac walls and valves have significantly increased the adoption of CT imaging systems in hospitals and clinical centers [1–3].

Kidney disease and renal failure represent major global public health challenges. The shortage of nephrologists and radiologists further exacerbates the burden associated with Chronic Kidney Disease (CKD). It is estimated that CKD affects more than 10% of the world's population and was ranked as the sixteenth leading cause of death in 2016. Moreover, it is projected to become the fifth leading cause of mortality by 2040. Consequently, early diagnosis of kidney-related disorders is crucial for preventing disease progression and reducing the risk of renal failure [4].

Motivated by this clinical necessity, automated kidney disease diagnosis based on medical image analysis has attracted considerable research attention. Noise reduction, image enhancement, tissue segmentation, and the development of highly accurate disease classification algorithms constitute the primary objectives of current research efforts. Although CT images generally provide superior image quality compared with ultrasound and Magnetic Resonance Imaging (MRI), image quality remains a critical factor affecting the performance of medical image processing techniques, particularly tissue segmentation algorithms. Tissue segmentation is widely recognized as the first and most important step in medical image analysis and computer-aided diagnosis. By accurately identifying tissue location, shape, and size, segmentation methods provide valuable information for detecting and characterizing pathological abnormalities [5].

However, the presence of inherent noise in CT images introduces significant challenges to automated kidney disease diagnosis. Appropriate preprocessing and enhancement techniques can substantially improve image quality, thereby facilitating more reliable disease detection when combined with efficient classification algorithms.

Numerous approaches have been proposed for image enhancement and noise suppression in CT imaging. Representative methods include weighted median filtering [6], the Savitzky–Golay filter [7], and wavelet-based denoising techniques [8]. In wavelet-based approaches, the original image is first decomposed into wavelet coefficients, which are subsequently modified according to a predefined strategy. The enhanced image is then reconstructed through the inverse wavelet transform. Several coefficient modification strategies have been proposed, including Bayesian estimation methods that exploit the statistical properties of wavelet coefficients. Nevertheless, thresholding remains the most widely adopted approach. Threshold-based denoising methods face two major challenges: 1) selecting an appropriate threshold value, which constitutes the most critical design parameter, and 2) addressing the varying statistical distributions of signal and noise across different wavelet scales. Various solutions have been proposed to address these challenges. While some studies apply identical processing strategies across all scales, others employ scale-dependent coefficient modification schemes [9], [10].

Following noise reduction, tissue segmentation constitutes another essential stage in kidney image analysis. Commonly used segmentation techniques include region-growing algorithms [1] and active contour models [2]. Due to the presence of image noise, region-growing methods often suffer from reduced reliability and segmentation accuracy [3]. Similarly, active contour models are not always suitable for kidney segmentation because kidney boundaries are frequently indistinct and exhibit weak contrast with surrounding tissues. Active Shape Models (ASMs) require manually selected landmark points to guide the segmentation process, which is both time-consuming and operator-dependent. Furthermore, fully automated landmark initialization techniques remain insufficiently mature [3].

A segmentation approach based on shape features and region-based energy functions was introduced in [11] and demonstrated satisfactory performance for MRI tissue segmentation. However, owing to the high level of intrinsic image noise, this method is less effective when applied to kidney CT images. Another group of researchers proposed a shape-based kidney modeling framework combined with a Gabor filter bank for texture feature extraction and segmentation [5]. In this approach, an average kidney shape model is first aligned with the test image. Subsequently, Gabor filters are applied to regions inside and outside the evolving boundary to extract texture features. An energy function is then formulated to maximize the similarity between the texture characteristics inside the boundary and a predefined kidney texture model while minimizing similarity outside the boundary. If the stopping criterion is not satisfied, gradient vectors and energy-transpose matrices are computed to update the boundary location. This iterative optimization process continues until the similarity measure reaches its maximum value. Although effective, the computational complexity of repeatedly applying Gabor filters and extracting texture features results in a relatively high processing time for each image slice [5].

Although semi-automatic segmentation methods have demonstrated promising performance in clinical kidney image analysis, fully automatic kidney segmentation remains challenging because of the considerable variability in kidney morphology and image intensity distributions [12]. Consequently, Deep Learning (DL) techniques have increasingly been adopted for automatic tissue segmentation and medical image analysis tasks [12–14].

In this study, a novel kidney disease classification framework is proposed based on image preprocessing, edge extraction, local statistical feature analysis, and DL using a Recurrent Neural Network (RNN). The proposed approach offers several advantages, including the elimination of complex handcrafted feature extraction procedures, high computational efficiency during the preprocessing stage, and excellent classification performance. The remainder of this paper is organized as follows. Section 2 presents the fundamentals of RNNs, with particular emphasis on Long Short-Term Memory (LSTM) architectures. Section 3 describes the proposed methodology in detail. Experimental results and performance evaluation are presented in Section 4. Finally, conclusions and future research directions are discussed in Section 5.

RNNs differ significantly from conventional Neural Network (NN) architectures in both their structural design and training mechanisms. Unlike feedforward NNs, RNNs incorporate recurrent connections that enable information from previous inputs to be fed back into the network during training. As a result, these networks possess an inherent short-term memory capability, allowing them to retain information from preceding inputs and utilize this information when processing subsequent data. This characteristic makes RNNs particularly suitable for modeling sequential and temporally dependent data.

Among DL architectures, RNNs are especially effective for processing time-series data and image sequences in which each sample may exhibit dependencies on preceding samples. However, standard RNN architectures suffer from several limitations. Due to their relatively simple recurrent structure, constrained hidden-state dynamics, and gradient-based weight update mechanisms, conventional RNNs often fail to effectively learn long-term dependencies when patterns span extended temporal intervals. As the temporal distance between relevant observations increases, the influence of earlier information gradually diminishes during training, ultimately approaching zero. Consequently, the network becomes unable to capture and retain long-range contextual information. This limitation, commonly known as the long-term dependency problem, represents one of the major weaknesses of traditional RNNs.

Various approaches have been proposed to address this issue, including alternative learning strategies and modifications to the recurrent architecture. Among these solutions, improved RNN structures have proven to be the most effective in simultaneously modeling both long-term dependencies and short-term variations. Owing to their ability to process sequential information efficiently, these architectures have become widely adopted in applications involving image analysis, video processing, speech recognition, and other tasks where the current output depends on both present and historical information.

One of the most successful extensions of conventional RNNs is the LSTM network. LSTM was specifically developed to overcome the long-term dependency limitations of standard recurrent networks by introducing an internal memory mechanism capable of preserving information over extended sequences. In a conventional RNN, the contribution of earlier observations gradually decreases as new inputs are processed, eventually becoming negligible. LSTM networks mitigate this problem through a gated memory structure that selectively stores, updates, and retrieves information. The architecture of an LSTM network is illustrated in

An LSTM cell consists of several functional components that regulate the flow of information through the network. In Fig. 1, σ denotes the activation function, while σg and σh represent the input and output activation functions, respectively. These functions are commonly implemented using the hyperbolic tangent (tanh) or logistic sigmoid functions, although the identity function may also be employed in certain configurations. The dashed connections shown in Fig. 1 correspond to weighted peephole connections. In peephole-based LSTM architectures, the input gate, output gate, and forget gate are influenced not only by the current input and previous hidden states but also by the previous internal cell states. This additional information increases the degrees of freedom of the model and enhances its capability to learn complex temporal relationships. Among the gating mechanisms, the forget gate plays a particularly important role. It determines which portions of the previous cell state should be retained and which should be discarded during the update process. By selectively preserving relevant historical information, the forget gate enables the network to capture long-term dependencies and maintain meaningful contextual information over extended sequences. Consequently, LSTM networks exhibit superior performance compared with conventional RNNs in applications involving long-range temporal correlations.

The overall framework of the proposed method is illustrated in Fig. 2.

As shown in Fig. 2, the first step involves removing the black margins surrounding the CT image. This cropping operation eliminates non-informative regions and reduces the image dimensions, thereby decreasing computational complexity and improving feature extraction efficiency.

Following the removal of the surrounding black borders, an asymmetric normalized histogram is computed from the remaining image region. The use of asymmetric histogram intervals is motivated by the non-uniform distribution of pixel intensities within kidney CT images. In particular, darker intensity regions generally exhibit a higher pixel density than brighter regions. Therefore, symmetric histogram partitioning may fail to provide sufficient discrimination between tissue characteristics. To address this issue, asymmetric intensity intervals are employed to achieve a more informative statistical representation of image content.

In addition to the intensity histogram, gradient-based information is extracted from the image. The image gradient highlights edge structures corresponding to anatomical boundaries and pathological abnormalities within the kidney region. Since most renal abnormalities manifest themselves through structural variations and edge-related characteristics, the statistical distribution of edge pixels can provide valuable information regarding the presence and type of disease. Consequently, gradient-based descriptors are incorporated into the feature extraction process to enhance discriminative capability.

To further capture local spatial information, each CT image is partitioned into one hundred non-overlapping regions of equal size, arranged in a 10 × 10 grid. Subsequently, an asymmetric normalized histogram is extracted from both the original image and its corresponding gradient image for each local region.

The histogram descriptor consists of eight bins. Therefore, for the original image, a feature matrix of size 100 × 8 is obtained. An additional feature matrix of identical dimensions is extracted from the gradient image. These two feature representations are then concatenated, resulting in a final feature matrix of size 100 × 16 for each CT image.

All feature values are normalized to the range [0,1], producing a compact and robust representation of both intensity and edge information. The resulting feature matrix serves as the input to the DL classifier.

The classification stage is performed using an LSTM recurrent neural network. The LSTM architecture is selected because of its ability to effectively model sequential dependencies within the extracted feature vectors while simultaneously preserving both local and long-range contextual information. The configuration of the employed LSTM network is summarized in Table 1.

The Softmax layer maps the input vector into a set of normalized values within the range of 0 to 1. This activation function provides a smooth and differentiable approximation of the maximum function, making it particularly suitable for gradient-based optimization algorithms. The outputs of the Softmax function represent a probability distribution over the target classes, such that the sum of all output probabilities is equal to one. Since the Softmax layer does not contain trainable parameters, its primary role is to transform the outputs of the preceding layer into normalized class probabilities within the interval [0,1]. The Softmax function is mathematically defined as follows:

where (xi) denotes the (i)-th element of the input vector (i.e., the output of the fully connected layer), and (yi) represents the corresponding Softmax output associated with that input.

The Cross-Entropy loss function is widely employed in classification problems where the network output represents a probabilistic value within the range of 0 to 1. This loss function quantifies the discrepancy between the predicted probability distribution and the ground-truth labels. As the predicted probability deviates from the actual class label, the Cross-Entropy value increases rapidly. Consequently, minimizing this loss function effectively drives the predicted probabilities toward the true target values, leading to improved classification performance.

When training a multi-class classifier whose final layer generates probabilistic outputs through the Softmax activation function, the Categorical Cross-Entropy (CCE) loss function is generally preferred due to its favorable optimization properties and probabilistic interpretation. CCE measures the divergence between the predicted probability distribution and the corresponding ground-truth labels, thereby encouraging the network to assign higher probabilities to the correct classes while suppressing incorrect predictions. As the predicted distribution deviates from the target distribution, the loss value increases rapidly, making it an effective objective function for multi-class classification problems [13], [14].

For a dataset containing (N) training samples and (C) target classes, the CCE loss is defined as:

LCCE = ∑ ∑ yij × log yij ̀ , (2)

i=1 j=1

where:

N: Number of training samples.

C: Number of classes.

yij: Ground-truth label.

yij ̀ : Predicted probability for class (j) of sample (i).

To optimize the network parameters, the Adaptive Moment Estimation (Adam) optimizer was employed. Adam combines the advantages of momentum-based optimization and adaptive learning-rate methods by maintaining exponentially weighted estimates of both the first and second moments of the gradients [15]. The parameter update equations are given by

mₜ = β₁mₜ₋₁ + (1 − β₁)gₜ, (3) vₜ = β₂vₜ₋₁ + (1 − β₂)gₜ², (4) m̂ ₜ = mₜ / (1 − β₁ᵗ), v̂ₜ = vₜ / (1 − β₂ᵗ), (5) θₜ₊₁ = θₜ − α × [ m̂ ₜ / (√v̂ₜ + ε) ], (6)

where gₜ denotes the gradient at iteration t, mₜ and vₜ are the first and second moment estimates, respectively, α is the learning rate, and ε is a small constant introduced to ensure numerical stability.

The combination of the Adam optimizer and CCE loss provides stable convergence, efficient parameter updates, and improved generalization performance. Consequently, this optimization strategy enables the proposed LSTM-based framework to achieve highly accurate classification results while maintaining robust training behavior across all kidney disease categories.

In this paper, a novel method based on image preprocessing, edge detection, and local statistical features of the image, combined with DL (RNN), is proposed for the detection and classification of kidney cancer. Given the critical importance of early detection of various cancers using medical images and timely treatment of patients, a fast and reasonably accurate classification of kidney conditions based on CT images can be highly beneficial. Considering the existing limitations, such as the shortage of specialist physicians in many regions and the high cost of clinical examinations, the proposed method, which eliminates redundant parts of the image, employs statistical histogram analysis in multiple adaptive windows, and simultaneously considers the effect of edges and the relationships between different regions, can detect kidney cancer with high accuracy and determine its type.

Due to the large number of pixels with low intensity in CT images, the histogram bin intervals were selected to be narrower near zero (low brightness) for both the original image and its gradient, and wider for higher brightness values. This adaptive binning strategy increases the discriminative power of the histogram and provides a more representative statistical description of the image, leading to improved classification accuracy for different types of kidney abnormalities.

The proposed method was implemented on images from the standard and publicly available Kaggle dataset, comprising four classes: cyst, normal, stone, and tumor. The classification accuracy achieved across these classes is 99.76%, which demonstrates a highly competitive performance compared to previous studies. Key advantages of the proposed approach include the elimination of complex feature extraction stages and its excellent computational efficiency.

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

All data are included in the text.

The authors declare no conflict of interest.