Renal cell carcinoma (RCC) is the most common type of kidney cancer, accounting for approximately 85-90% of kidney malignancies. Once a tumor is removed or biopsied, pathologists examine the tissue under a microscope to assign a tumor grade - a measure of how abnormal the cancer cells appear and how aggressively the tumor is likely to behave. Higher grade tumors tend to grow faster and carry worse prognoses.
The current standard grading system, known as the WHO/ISUP grading system, classifies RCC into grades 1 through 4 based on the appearance of cell nuclei under microscopy. However, manual grading is inherently subjective. Two experienced pathologists examining the same slide may reach different conclusions, particularly for tumors that fall on the border between grades. This inter-observer variability can affect treatment decisions and prognostic accuracy.
Automating tumor grading using artificial intelligence could reduce this variability, speed up diagnosis, and potentially reveal subtle visual patterns that the human eye consistently misses. This paper presents EAT-Net (Efficient Attention Transformer Network), a deep learning model designed to automatically classify RCC tumor grades from histopathological images.
EAT-Net uses a dual-stream architecture that runs two different types of neural networks in parallel and combines their outputs. The first stream is EfficientNetB0, a convolutional neural network (CNN) known for achieving high accuracy with relatively few computational parameters. CNNs are specialized for detecting local spatial features - fine-grained details like the shape and texture of individual cell nuclei.
The second stream is a custom Vision Transformer (ViT), which processes images by breaking them into patches and learning how different regions of the image relate to one another through a mechanism called self-attention. Unlike CNNs, which look at local neighborhoods of pixels, the ViT captures global contextual information - for example, how the spatial arrangement of cells across the entire slide relates to tumor grade. The custom ViT was designed to be lightweight (about 4 million parameters, compared to 86 million for standard ViT architectures), maintaining efficiency without sacrificing performance.
Both streams process the same input image simultaneously. The outputs are concatenated and used to predict the tumor grade (grades 0 through 4). Squeeze-and-Excitation (SE) modules within the EfficientNet stream further improve performance by allowing the network to emphasize the most informative image channels and suppress less relevant ones. The model was trained and tested on two publicly available RCC histopathology datasets: KMC-RENAL and RCCG-Net.
EAT-Net achieved an overall accuracy of 92.25%, with precision of 92.15%, recall of 92.12%, and an F1-score of 92.25% on the external test dataset (RCCG-Net). These results outperformed all comparison models tested in the study, including ResNet (74.64% accuracy), standard ViT (82.39%), NASNet (80.28%), and the previous best-performing specialized model EFF-Net (91.90%).
Ablation studies - experiments that methodically remove components of the model to test their contribution - confirmed that both streams are necessary. Using EfficientNetB0 alone achieved 92.96% accuracy on the internal test set, while adding the ViT stream improved consistency across all metrics in the combined evaluation on the external dataset. The complementary nature of local feature extraction (EfficientNet) and global context modeling (ViT) proved beneficial for handling the variety of patterns encountered across all five tumor grades.
Grad-CAM visualizations (a technique that highlights which image regions most influenced the model's prediction) showed that EAT-Net consistently focused on diagnostically relevant areas such as cell nuclei with meaningful morphological features. In contrast, the EfficientNet-only model more frequently attended to irrelevant background regions when making incorrect classifications - demonstrating that the ViT stream's global attention mechanism helps guide the model toward clinically meaningful tissue regions.
EAT-Net processes a single histopathology image in approximately 17.95 milliseconds, making it fast enough for real-time use in a clinical pathology workflow. With only 12.1 million parameters and 1.99 GFLOPS of computational cost, it is also lightweight enough to run on standard hospital computing infrastructure without requiring specialized high-performance hardware.
The model is explicitly designed as an assistant for pathologists, not a replacement. By providing rapid, consistent grade predictions alongside Grad-CAM visualizations showing why the model reached its conclusion, EAT-Net gives pathologists an objective second opinion. In practice, a pathologist could review the model's prediction and highlighted regions, then confirm or override based on their expert judgment - combining the speed and consistency of AI with the interpretive expertise of a trained clinician.
Automated grading tools like EAT-Net could be particularly valuable in settings with limited pathologist availability or high workload. In lower-resource clinical environments where specialist pathology expertise is scarce, an AI-based grading assistant could help ensure more consistent tumor characterization across all patients, regardless of local expert availability.
EAT-Net demonstrates that combining local feature extraction from CNNs with global contextual reasoning from transformers produces a superior model for kidney cancer grading compared to either approach alone. The result is a lightweight, accurate, and interpretable tool that achieves state-of-the-art performance while remaining practical for clinical deployment.
For patients, consistent and accurate tumor grading matters because it directly influences treatment decisions. Misclassifying a high-grade tumor as low-grade could lead to under-treatment, while over-grading could lead to unnecessarily aggressive interventions. An AI tool that reduces grading variability between pathologists helps ensure patients receive the treatment appropriate to their actual tumor biology.
The authors acknowledge limitations including the use of only two publicly available datasets (which may not represent all clinical diversity), potential class imbalance affecting performance on higher-grade tumors, and the need for external validation in large multi-institutional prospective studies. Future work may also explore incorporating molecular features alongside histopathology images to further improve grading accuracy in line with evolving WHO classification standards.