Accurate grading of kidney cancer is essential for treatment planning, as higher-grade tumors are associated with more aggressive behavior and poorer outcomes. Traditional grading relies on pathologist examination of tissue samples, which is time-consuming and subject to inter-observer variability. Automated AI-based grading from imaging could standardize this process and reduce diagnostic delays.
The Internet of Medical Things (IoMT) refers to the network of connected medical devices, sensors, and systems that generate and transmit healthcare data. While IoMT enables new capabilities for remote monitoring and AI-assisted diagnostics, it also introduces significant security vulnerabilities: sensitive medical imaging data transmitted over networks can be intercepted, altered, or stolen by malicious actors.
This study combined transfer learning using the AlexNet convolutional neural network architecture for kidney cancer grade prediction with a blockchain-based security framework to protect model weights and patient data. The goal was to demonstrate that high-accuracy AI-based grading can be achieved while maintaining robust data security in an IoMT environment.
AlexNet is a classic deep convolutional neural network architecture that achieved breakthrough performance in the ImageNet Large Scale Visual Recognition Challenge in 2012, demonstrating the power of deep CNNs for image classification. Despite being over a decade old, AlexNet remains a widely used transfer learning base model because its relatively compact architecture is efficient to fine-tune on smaller datasets.
Transfer learning involves taking a neural network pre-trained on a large dataset (in this case, ImageNet, which contains millions of natural images across 1,000 categories) and adapting it to a new task by retraining the final layers on task-specific data. The early convolutional layers, which have learned general visual features such as edges, textures, and shapes, are typically frozen or only lightly updated, while the classification head is replaced and trained on the target dataset.
The study applied transfer learning to classify kidney cancer images into three grade categories: grade 0 (no cancer), grade 1 (low-grade), and grade 2 (high-grade). Data augmentation was used to expand the dataset to 3,300 images, addressing the limited size of the original imaging dataset and reducing the risk of overfitting during fine-tuning.
The study compared three optimization algorithms commonly used for training deep neural networks: Stochastic Gradient Descent with Momentum (SGDM), ADAM (Adaptive Moment Estimation), and RMSProp (Root Mean Square Propagation). Each optimizer uses a different strategy for adjusting model weights during training based on the gradient of the loss function.
SGDM adds a momentum term to standard gradient descent, allowing the optimizer to accelerate in directions of consistent gradient and dampen oscillations. ADAM combines momentum with adaptive learning rates for each parameter, generally converging faster and more reliably on deep learning tasks. RMSProp also uses adaptive learning rates but without the momentum term, making it more suitable for recurrent problems and noisy gradients.
All three optimizers were applied to the same AlexNet transfer learning setup, allowing direct comparison of their impact on training and test accuracy for the kidney cancer grading task. Understanding optimizer choice is practically important because it affects both convergence speed and final model accuracy.
All three optimizers produced extremely high classification accuracy, demonstrating the effectiveness of transfer learning even with a relatively small dataset. SGDM achieved the best performance with 99.8% training accuracy and 99.2% testing accuracy. ADAM followed with 99.0% training accuracy and 98.3% testing accuracy. RMSProp achieved 98.98% training accuracy and 98.18% testing accuracy.
The near-perfect testing accuracy across all optimizers is a strong finding, though it should be interpreted with caution. The 3,300-image dataset was generated through augmentation from a smaller original set, which means augmented images are not fully independent of their source images. If augmented images from the same original scan appear in both training and test splits, test accuracy may be inflated relative to performance on genuinely new patients.
Nevertheless, the consistency of high performance across three different optimizers suggests that the transfer learning approach is robust to the choice of optimization algorithm for this specific task. The ordering of optimizer performance (SGDM greater than ADAM greater than RMSProp) is consistent with some prior literature on transfer learning for medical image classification, though results vary by dataset and architecture.
The blockchain component of this study addresses the security requirements of deploying AI models in an IoMT environment. Blockchain is a distributed ledger technology in which data is organized into cryptographically linked blocks, making it tamper-evident: any attempt to alter previously recorded data would require altering all subsequent blocks, a computationally infeasible task in a properly configured blockchain network.
The study used private blockchain clouds, where a restricted set of authorized nodes (representing participating healthcare institutions or devices) maintain the distributed ledger. Both AI model weights and patient imaging data were secured using blockchain-based access controls, ensuring that only authorized parties could access or modify these assets.
This approach addresses a real and growing concern in medical AI: as models are deployed across networks of IoMT devices, the integrity of both the model and the data it processes must be guaranteed. A compromised model could misclassify tumors, while compromised patient data could violate privacy regulations. The blockchain framework provides cryptographic guarantees of integrity that traditional centralized security systems cannot match.
The combination of high-accuracy transfer learning and robust security infrastructure demonstrated in this study addresses two of the most important barriers to real-world AI deployment in oncology: performance and security. A model that achieves 99% accuracy but can be compromised is clinically unacceptable, just as a secure model that performs at chance is worthless.
The IoMT-blockchain framework could enable distributed deployment of kidney cancer grading AI across multiple clinical sites, with each site contributing to model use while maintaining local data sovereignty. This aligns with the increasing emphasis on privacy-preserving distributed AI in healthcare, as embodied in frameworks such as federated learning.
Future work should validate this approach on larger and more diverse patient datasets, assess performance on genuinely independent test sets without augmentation overlap, and evaluate the practical feasibility and computational overhead of blockchain integration in real clinical environments. Regulatory frameworks for AI tools secured through blockchain also remain to be fully defined.