This study introduces a new AI-powered system called EDTL-PCGRCC, designed to automatically classify and grade kidney cancer using images of tissue samples. The system achieved over 97% accuracy in testing, which is substantially better than other leading deep learning methods.
Accurately grading kidney cancer, meaning determining how aggressive the tumor cells look under a microscope, is critical for planning treatment. The higher the grade, the more aggressive the cancer tends to be. Currently, grading is done manually by pathologists and can vary between different specialists. An AI system that performs this task consistently and accurately could reduce errors and speed up diagnosis.
Published in Scientific Reports in 2025, this research advances the field of computational pathology and demonstrates how carefully combining multiple AI components can produce a system that outperforms both human benchmarks and previous AI models.
Renal cell carcinoma (RCC) is the most common type of kidney cancer and comes in several subtypes with different behaviors. Within each subtype, tumor grade reflects how abnormal the cancer cells appear, with higher grades corresponding to more aggressive tumors that are more likely to spread.
Traditional grading relies on a pathologist examining tissue samples stained with dyes and viewed under a microscope. This process is time-consuming, requires expert training, and can produce different results when the same slide is reviewed by different specialists, a problem called inter-observer variability. AI systems could make grading faster and more consistent across hospitals and regions.
Previous AI approaches to RCC grading have often used standard deep learning architectures that were not specifically optimized for pathology images. The researchers behind this study identified several weaknesses in existing models and designed a pipeline to address them directly.
The EDTL-PCGRCC system processes tissue images through four sequential stages. In the first stage, a mathematical technique called a Wiener Filter (WF) removes noise and artifacts from the images, improving their quality before the AI begins analyzing them. Clean images lead to more reliable feature detection in later stages.
In the second stage, a modified version of MobileNetV2 extracts meaningful features from the cleaned images. MobileNetV2 is a lightweight but powerful neural network originally designed for mobile devices. The researchers improved it by replacing its standard activation function with SELU (Scaled Exponential Linear Unit), which helps the network learn more efficiently and avoid common training problems like vanishing gradients.
The third stage uses an Elman Neural Network (ENN) to classify the images into one of five categories: Grade 0 (normal tissue) through Grade 4 (highly aggressive cancer). ENN is a type of recurrent neural network with a memory-like component, making it well-suited to capturing sequential patterns in data. The fourth and final stage uses an algorithm called Improved Artificial Ecosystem Optimization (IAEO) to automatically tune the ENN's settings for maximum accuracy.
The system was trained and tested on a dataset of 3,000 kidney cancer tissue images across five classes: Grade 0 (normal), Grade 1, Grade 2, Grade 3, and Grade 4. Each class contained exactly 600 images, ensuring balanced representation so the AI would not develop a bias toward more common grades.
The researchers tested the system under two conditions: using 80% of the data for training and 20% for testing, and using 70% for training and 30% for testing. This approach, called cross-validation, checks whether the system works consistently regardless of which specific images it was trained on.
The SELU activation function used in the modified MobileNetV2 is mathematically designed to keep neuron outputs in a healthy range during training, a property called self-normalizing. This means the network trains more stably and often converges faster than with older activation functions like ReLU6, which the original MobileNetV2 used.
With the 80/20 training/testing split, EDTL-PCGRCC achieved an accuracy of 97.20%. With the 70/30 split, it achieved 96.98%. Both results are remarkably high for a five-class classification problem involving complex tissue images.
The system clearly outperformed other widely-used deep learning models tested on the same dataset. For comparison, ResNet34 achieved 90.32% and DenseNet121 achieved 90.29%. These are well-established, respected architectures, and the 6-7 percentage point improvement achieved by EDTL-PCGRCC is clinically meaningful.
The system also performed well on precision, recall, and F1 score, which measure different aspects of classification accuracy beyond simple overall correctness. High performance on all these metrics together suggests the system is not just getting easy cases right but is reliably distinguishing between grades that are genuinely difficult to tell apart.
Accurate grade assignment directly affects treatment decisions. Lower-grade tumors may be managed with active surveillance or less aggressive surgery, while higher-grade tumors typically require more extensive treatment and closer follow-up. An AI system that correctly grades tumors with over 97% accuracy could help ensure patients receive the right level of care.
The system's consistency is also important. Unlike human pathologists, an AI does not get tired, distracted, or influenced by subjective factors. This means a system like EDTL-PCGRCC could deliver the same quality of grading at 3 in the morning as at 9 in the morning, and in a hospital with limited pathology expertise as well as in a major academic center.
Before clinical adoption, the system would need to be tested on larger, more diverse datasets from different hospitals and patient populations, and validated in prospective clinical trials. However, the results presented here are a strong foundation for that next phase of development.
EDTL-PCGRCC demonstrates that combining multiple AI components, each optimized for a specific part of the problem, can produce a system that significantly outperforms single-model approaches. The choice to use SELU activation in MobileNetV2 and to tune the classifier with an evolutionary algorithm shows the value of thoughtful architectural decisions rather than simply using standard off-the-shelf models.
The five-class grading capability is particularly noteworthy. Many existing AI systems for kidney cancer focus on binary tasks, such as benign versus malignant. The ability to distinguish between multiple grades adds clinical nuance and makes the system more useful in real diagnostic settings.
The researchers acknowledge that future work should include testing on histopathology images from multiple institutions, incorporating a wider range of kidney cancer subtypes, and exploring how the system performs when integrated into a real pathology workflow. These are important next steps toward translating this promising research into a tool that genuinely helps patients.