The scale of the problem. Lung cancer is the leading cause of cancer-related deaths worldwide, responsible for approximately 18% of all cancer fatalities. Early detection dramatically improves survival odds, but deploying reliable AI diagnostic tools in clinical settings faces persistent structural barriers that go beyond simply building an accurate model.
Computational bottlenecks with large datasets. Modern lung cancer detection relies on CT scans, X-rays, and other imaging data at massive scale. Traditional AI systems struggle to process these large datasets within the time constraints of clinical workflows, leading to delays that undermine the practical utility of even highly accurate models.
Privacy barriers to data sharing. Hospitals and healthcare institutions hold sensitive patient data governed by strict regulations such as HIPAA in the United States and GDPR in Europe. These regulations prevent institutions from simply pooling their data to train better AI models, leading to fragmented datasets and models that may not generalize well across different patient populations or clinical settings.
The black box problem. Even when AI models perform well statistically, clinicians are reluctant to rely on systems they cannot understand. If a model predicts lung cancer without explaining which image features drove that prediction, a physician has no way to verify whether the AI's reasoning aligns with medical knowledge - a fundamental trust barrier to clinical adoption. This study addresses all four of these challenges simultaneously with a unified framework.
MapReduce for computational scale. MapReduce is a distributed computing framework that splits large datasets into smaller chunks, processes them in parallel across multiple machines, then combines the results. Applied to medical imaging, it eliminates the processing bottlenecks that occur when a single machine tries to handle thousands of CT scan images, enabling faster model training and faster diagnostic support.
Private blockchain for data security. Blockchain technology creates an immutable, tamper-proof ledger of transactions. In this context, patient data records and model transactions are stored as cryptographically secured blocks distributed across a decentralized network of hospital nodes. No single entity controls or can silently alter the data, and smart contracts automate compliance with privacy policies - aligning with HIPAA and GDPR requirements.
Federated learning for collaborative privacy-preserving training. Federated Learning (FL) allows multiple hospitals to collaboratively train a shared AI model without ever transferring raw patient data off their local servers. Each institution trains the model locally on its own data, then shares only the model parameters (weights and biases) - not the underlying patient records - with a central server that aggregates them into an improved global model.
Explainable AI for clinical transparency. Explainable Artificial Intelligence (XAI) techniques produce human-readable explanations of what features drove a model's prediction. Rather than simply outputting "cancer" or "no cancer," the system highlights which specific regions of the CT scan image were most influential, enabling clinicians to verify that the AI is focusing on medically relevant structures rather than imaging artifacts or irrelevant patterns.
The dataset used. The model was trained and tested on a publicly available lung cancer dataset from Kaggle, comprising CT scans and X-ray images from 1,097 patients total - 120 benign cases, 561 malignant cases, and 416 normal cases. The data was split 70% for training and 30% for testing.
Preprocessing for image quality. Before feeding images to any AI model, they underwent several preprocessing steps designed to improve quality and ensure consistency. Normalization scaled pixel intensities to a uniform range, making images from different scanners more comparable. Noise reduction applied Gaussian blur (smoothing overall) and median filtering (better for preserving sharp edges) to remove imaging artifacts.
Edge detection and feature enhancement. Two edge detection techniques - Canny edge detection for fine structural detail and Sobel edge detection for broader features - were applied to highlight the boundaries of lung regions, nodules, and tissue abnormalities. These steps help deep learning models more easily identify clinically relevant structures within the image.
Color space transformations. Although CT scans are grayscale, transforming images into different color spaces can reveal additional information. The HSV (Hue, Saturation, Value) space allows finer control over brightness and contrast differences, while the LAB space separates lightness from color information, making it effective for texture analysis and contrast enhancement - both relevant for identifying subtle lung abnormalities.
MapReduce processing pipeline. Preprocessed image data is divided into chunks and passed through three phases: the Map phase transforms each chunk into intermediate key-value pairs (for example, feature label plus pixel intensity); the Shuffle phase sorts and groups results by key; and the Reduce phase aggregates grouped data using operations like averaging. This parallel architecture dramatically accelerates the feature extraction and training process for large imaging datasets.
Secure blockchain storage of processed data. After MapReduce processing, training and testing datasets are stored on the private blockchain layer. Each data transaction is converted into a secure cryptographic block, distributed across decentralized hospital databases, and validated by trusted peer nodes. This ensures that data cannot be silently tampered with during the collaborative training process.
Federated learning aggregation. Individual hospitals train local deep learning models on their own data. When performance meets defined criteria, each hospital sends only its trained model weights and biases - not patient records - to a global cloud server. The server aggregates these local model parameters into an improved global model, then redistributes it back to each hospital. This cycle continues iteratively, progressively improving the shared model without centralizing sensitive patient data.
XAI explanation methods applied to the global model. After federated training, two XAI methods are applied. LIME (Local Interpretable Model-agnostic Explanations) creates a simpler surrogate model that approximates the complex AI model's behavior near a specific input, identifying which image regions (superpixels) most influenced the prediction. Grad-CAM (Gradient-weighted Class Activation Mapping) uses gradient information from the model's convolutional layers to generate color-coded heatmaps showing which image regions were most important for the final classification decision.
Four local models evaluated. Four deep learning architectures were trained and tested as local models within the federated framework: a standard CNN, EfficientNetB0, InceptionV3, and DenseNet121. Each was assessed on training and testing accuracy alongside miss rate - the percentage of cases incorrectly classified.
CNN outperforms the specialized architectures. Despite its simpler architecture compared to more recent designs, the CNN model achieved the highest testing accuracy at 98.21% with a miss rate of just 1.79%. EfficientNetB0 achieved 90.30%, InceptionV3 reached 93.31%, and DenseNet121 achieved 94.64%. The CNN's strong performance here may reflect the dataset's characteristics and the beneficial effect of the preprocessing pipeline.
The CNN becomes the global federated model. Based on its superior accuracy and low miss rate, the CNN was selected as the global model to be aggregated via federated learning. This means that insights from all participating institutions' local models are combined and weighted to produce a final shared model that generalizes across the diverse data from multiple hospitals.
XAI visualizations validate model focus. LIME analysis produced highlighted image regions (shown in yellow in the paper's figures) identifying the most influential areas driving lung cancer predictions. Grad-CAM generated heat maps where warmer colors (red, yellow) indicated regions of highest importance to the model's decision - typically corresponding to nodules or tissue abnormalities - while cooler colors indicated less relevant regions. These visualizations allow clinicians to confirm that the AI is examining medically appropriate areas of the scan.
Outperforming a wide field of competitors. Comparative analysis against 16 prior published lung cancer detection models shows the proposed framework achieving 98.21% accuracy - among the highest reported in the literature. Most comparison models achieved between 80% and 95% accuracy, with only one comparable model (a 2D CNN approach by Sathish et al.) achieving a similar 98.4% accuracy.
Unique combination of capabilities. Critically, the proposed system is the only one in the comparison table that combines all four key features: MapReduce computational efficiency, private blockchain security, federated learning, and XAI. Every prior model lacked at least one of these capabilities, meaning they traded off accuracy, privacy, scalability, or interpretability against each other. This framework demonstrates that high accuracy can be achieved without these trade-offs.
Unique value of the integrated approach. The paper argues that the real innovation is not just the accuracy number but the ability to achieve it in a privacy-preserving, scalable, and interpretable manner that meets real-world clinical and regulatory requirements. A model that achieves 98% accuracy but cannot be deployed at scale across multiple hospitals, or cannot be trusted by clinicians, delivers less real-world value than a slightly less accurate but deployable and interpretable system.
Aligning with regulatory standards. By design, the framework's architecture aligns with HIPAA and GDPR requirements: raw patient data never leaves local hospital systems, the blockchain provides audit trails and tamper-proof storage, and XAI ensures that AI decisions can be reviewed and explained to patients and regulators - all prerequisites for clinical AI adoption in practice.
Single dataset reliance. A significant limitation of the current study is that it relies on a single publicly available dataset from Kaggle. While the dataset includes over 1,000 cases across benign, malignant, and normal categories, it may not fully represent the diversity of patients encountered across different clinical settings, scanner types, imaging protocols, and global populations. Dataset-specific bias could make the model less reliable when deployed at other institutions.
Adversarial testing not yet performed. The framework has not been tested against adversarial attacks - deliberate manipulations of input images designed to fool AI models into making incorrect predictions. In high-stakes medical applications, adversarial robustness testing is important for validating security and reliability before deployment.
Computational overhead of integration. Running four complex technological systems simultaneously - MapReduce, blockchain, federated learning, and XAI - introduces additional computational cost compared to simpler standalone models. For healthcare settings with limited IT infrastructure, this overhead may present a practical barrier to deployment, particularly in low-resource or rural clinical environments.
Planned enhancements. Future research directions include multi-center validation using real hospital data, optimizing federated learning communication protocols to reduce overhead, incorporating expert-annotated overlays into XAI outputs to improve their clinical relevance, conducting formal ablation studies to quantify the individual contribution of each component, and adding statistical validation with additional metrics such as sensitivity, specificity, and F1-score.
More than an accuracy benchmark. The 98.21% accuracy achieved by this framework is noteworthy, but the paper's more significant contribution is demonstrating that high diagnostic performance can coexist with privacy preservation, collaborative learning, and clinical transparency. This challenges the common assumption that these goals necessarily trade off against one another.
Addressing clinician trust systematically. The XAI component directly confronts the most fundamental barrier to AI adoption in clinical medicine: the inability to understand and verify AI decisions. By highlighting which regions of a CT scan drove a prediction, clinicians can evaluate whether the AI is reasoning medically soundly - a prerequisite for the AI to be trusted as a decision-support tool rather than a black box to be bypassed.
A scalable model for healthcare AI governance. The framework's design - local training with only parameter sharing, blockchain-secured data management, and interpretable outputs - could serve as a template for responsible AI deployment in other high-stakes clinical areas beyond lung cancer detection, wherever data privacy, regulatory compliance, and clinical trust are essential requirements.
The path to real-world deployment. Realizing the clinical potential of this framework will require validation across multiple hospitals with diverse patient populations, integration with existing electronic health record systems, and engagement with clinical governance bodies to ensure the system meets regulatory standards for medical device software. The technical foundations demonstrated here provide a strong starting point for that translational journey.