AI-POWERED RENAL DIAGNOSIS

Introduction

Kidney disease remains a major global public health challenge and continues to rise in prevalence despite significant advancements in preventive and therapeutic strategies¹. Early diagnosis is essential to slowing the progression of Chronic Kidney Disease (CKD) and mitigating its potentially-severe consequences. Current estimates indicate that more than 10% of the global population are affected by CKD, which is expected to become the sixth leading cause of death worldwide by 2040². Alarmingly, CKD currently affects approximately 850 million individuals, more than double the global prevalence of diabetes and twenty times that of cancer³. This growing burden is primarily attributed to the increasing incidence of common risk factors, such as diabetes mellitus and obesity, which contribute substantially to the development and progression of CKD.

Research Problem

Renal diseases, including tumors, cysts, and kidney stones, present substantial diagnostic and therapeutic challenges on a global scale. Timely and accurate identification of these conditions is essential for preventing progression to renal failure and for reducing associated morbidity and mortality. However, current diagnostic workflows are heavily dependent on radiological expertise, which remains critically limited, especially in low-resource and underserved healthcare settings. Although conventional machine learning techniques have demonstrated utility in certain diagnostic applications, they frequently encounter significant limitations. These include poor scalability, inadequate support for multiscale feature extraction, and limited capacity to effectively integrate complex spatial and channel-wise information from medical images. Such constraints compromise their reliability and applicability for comprehensive renal disease diagnosis in real-world clinical environments. Given these challenges, there is an urgent need for advanced computational frameworks capable of delivering high diagnostic precision, while maintaining robustness and adaptability across diverse imaging datasets. A successful solution must effectively leverage multiscale and attention-based mechanisms to overcome existing limitations in feature learning, thereby supporting automated, scalable, and clinically-viable kidney disease detection.

Research Motivation

The rising global burden of renal diseases, coupled with a persistent shortage of trained radiologists and nephrologists, underscores the critical need for automated, intelligent diagnostic systems. Recent advancements in DL, particularly the development of attention mechanisms, present a compelling opportunity to address these challenges by enabling more effective feature extraction, pattern recognition, and clinical decision support. While CNNs have shown considerable promise in medical image analysis, many existing approaches fall short in capturing multiscale contextual information or in dynamically emphasizing diagnostically relevant regions within complex imaging data. These limitations often lead to diminished classification accuracy and restricted generalization in real-world clinical scenarios. This study is driven by the need to overcome such shortcomings through the design of a robust and adaptive attention mechanism that not only enhances diagnostic precision but also improves model interpretability and computational efficiency. By bridging the gap between clinical requirements and current technological capabilities, the proposed approach aims to support healthcare professionals in delivering timely, accurate, and scalable diagnoses of renal diseases.

Main Contribution

This study presents the PCSA, a novel attention mechanism specifically developed to enhance multiscale feature integration and adaptive learning for renal disease diagnosis. The PCSA block leverages pyramidal multiscale convolution to extract both spatial and channel attention weights, enabling precise feature reconstruction and targeted focus on diagnostically significant image regions. This design substantially improves the model ability to capture and discriminate intricate anatomical patterns commonly found in renal pathologies. The PCSA block is implemented as a modular, plug-and-play component, facilitating seamless integration into a variety of CNN backbone architectures. To demonstrate its effectiveness, the PCSA block was incorporated into the ResNet architecture, resulting in the development of the PCSANet, an enhanced model that achieves superior diagnostic accuracy and computational efficiency compared to existing state-of-the-art methods, while maintaining a reduced parameter footprint. Moreover, the flexibility of the PCSA module allows it to be readily adapted for broader medical imaging applications beyond renal disease classification. The primary contributions of this paper are summarized as follows:

• We introduce a novel dual-attention mechanism, the PCSA, that concurrently captures and integrates multi-scale spatial and channel information to enhance feature representation.
• The PCSA is designed as a lightweight, modular component that can be easily embedded into various backbone architectures to improve model performance.
• By integrating the PCSA into the ResNet architecture, we construct the PCSANet, a set of models that effectively learn complex multiscale features, while significantly reducing the number of parameters.
• Experimental results demonstrate that PCSANet models outperform several state-of-the-art methods in terms of diagnostic accuracy, adaptability, and efficiency through precise and adaptive channel-wise weight-ing.

Related Work

Deep learning (DL) and machine learning (ML) algorithms have shown considerable effectiveness in the prediction and diagnosis of various complex diseases. In recent years, the early detection of chronic conditions, particularly CKD, has garnered increased attention from both clinicians and researchers. Numerous emerging studies have explored the application of DL techniques for improving diagnostic accuracy of CKD and related renal disorders. This section provides a comprehensive review of recent contributions in this domain, highlighting both the advancements and the persisting limitations of existing approaches. In⁵, the authors utilized the publicly-available CT kidney dataset, Normal-Cyst-Tumor-Stone, comprising 12,446 annotated CT images to evaluate the performance of six ML models. These included three advanced Vision Transformer (ViT) architectures, Swin Transformer, Compact Convolutional Transformer (CCT), and External Attention Network (EANet), as well as three DL models, namely ResNet50, Inception V3, and VGG16. Among these, the Swin Transformer exhibited the highest classification performance, achieving a maximum accuracy of 99.30%, demonstrating its effectiveness in renal image analysis. Another approach¹⁴ presented a Deep Neural Network (DNN) framework for the early detection and prediction of CKD. This model depends on Recursive Feature Elimination (RFE) to identify key clinical predictors, including packed cell volume, specific gravity, hemoglobin levels, red blood cell count, serum creatinine, albumin, and hypertension. Using the UCI kidney dataset, the DNN model performance was compared with those of traditional classifiers such as K-Nearest Neighbors (KNN), Support Vector Machines (SVM), Logistic Regression (LR), Naive Bayes (NB), and Random Forest (RF). Remarkably, the DNN model achieved a perfect classification accuracy of 100%, highlighting its diagnostic potential in structured clinical datasets.

Proposed Methodology

This section outlines the architectural framework and theoretical foundation of the proposed PCSA-Net, a DL model designed for efficient and accurate renal disease classification using CT images. The development of the PCSA module is motivated by the limitations observed in existing attention mechanisms such as Squeeze-and-Excitation Networks (SE-Nets), Convolutional Block Attention Module (CBAM), and EPSA-Net. These previous methods either focus exclusively on channel or spatial attention or inadequately capture multiscale contextual information, which is critical for identifying subtle and heterogeneous patterns in medical images. To address these shortcomings, the PCSA module is designed with a dual-branch architecture that integrates both spatial and channel attention within a pyramid convolutional framework. This configuration enables adaptive emphasis on both high-level structural patterns and localized diagnostic cues across a range of receptive fields. Theoretically, this design leverages the complementary strengths of parallel attention pathways and multiscale feature fusion, enhancing the model ability to learn discriminative representations, while minimizing feature redundancy. The proposed PCSA module is seamlessly integrated into a ResNet backbone, forming the core of PCSA-Net. This integration enhances the network capacity to selectively focus on salient features across spatial and channel dimensions, thus improving the robustness and interpretability of deep feature learning in complex medical images. We begin by presenting the theoretical rationale for embedding attention mechanisms into CNN backbones. These mechanisms are essential for guiding the model focus toward clinically-relevant features, ultimately leading to improvements in both computational efficiency and classification accuracy. Specifically, we analyze the complementary roles of Channel Attention (CA) and Spatial Attention (SA). Channel Attention (CA) prioritizes feature channels that contain highly-informative content, enhancing the discriminative power of the model. Spatial Attention (SA) highlights spatial regions within the feature maps that are most indicative of the target pathology.

Results and Discussions

This section presents a comprehensive performance analysis of the proposed PCSANet model, based on extensive experiments conducted on a multiclass kidney CT image dataset. The classification capability of the model is evaluated using the top-1 accuracy metric, which quantifies the proportion of test samples for which the predicted class label matches the ground truth. To ensure the statistical reliability and robustness of the evaluation, the experiments were repeated multiple times under identical conditions, thereby accounting for inherent variability in DL model training. The final reported accuracy represents the average performance across these independent trials, offering a more stable and generalizable measure of the model diagnostic efficacy. The results ensure the superiority of PCSANet in capturing subtle renal pathologies and distinguishing between the four clinical categories: normal, cyst, stone, and tumor. The quantitative assessment of the proposed PCSANet model was conducted using a comprehensive set of standard evaluation metrics, including accuracy ³⁴, precision, recall³⁵, F1-score²⁴, support, macro average, and weighted average³⁶. The experimental findings, summarized in Tables 3, 4 and 5, demonstrate the effectiveness of the PCSANet model with different integration strategies. Among these, the standard integration approach consistently delivered superior performance across all evaluated metrics, including accuracy, precision, recall, and F1-score. Table 3 presents the results for the standard integration of the PCSA module within the ResNet architecture. This configuration achieved perfect scores (1.00) across all classes and metrics, indicating a highly discriminative and well-generalized model.

Conclusion & Future Works

This study introduced the PCSA mechanism, an advanced attention framework designed to improve feature extraction in convolutional neural networks. The PCSA leverages enhanced pyramid multiscale convolution to capture rich feature representations across varying receptive fields and generates both channel and spatial attention weights to effectively recombine multiscale features. This mechanism optimizes spatial information flow and improves the network focus on relevant regions. By replacing standard convolutions with dilated pyramid group convolutions, the PCSA reduces computational overhead, while pointwise convolution is employed to compensate for potential information loss inherent in grouped operations. As a lightweight and modular attention block, the PCSA is compatible with a wide range of deep learning architectures and can be integrated as a plug-and-play component. In this work, the integration of the PCSA into ResNet architectures led to the development of the PCSANet, a novel model that demonstrated superior classification accuracy and robustness compared to several state-of-the-art methods across a large-scale CT kidney dataset. Looking forward, future research will focus on extending the application of the PCSA to additional computer vision tasks such as object detection, semantic segmentation, and lesion localization, particularly within medical imaging domains. To further validate clinical utility, we plan to conduct external evaluations using multi-institutional datasets that incorporate heterogeneous patient demographics and imaging protocols. Additionally, to facilitate deployment in resource-constrained environments, we will explore model compression techniques and design lightweight variants of the PCSA module. These future directions aim to broaden the applicability, scalability, and practical adoption of the proposed attention mechanism in real-world healthcare and AI systems.