Renal cell carcinoma (RCC) is the most common type of kidney cancer, accounting for approximately 90% of all kidney cancers. Kidney tumors often grow silently without symptoms in their early stages, meaning that by the time most patients are diagnosed, the cancer may have already progressed. Early and accurate detection is one of the most important factors in improving survival rates.
Computed tomography (CT) scans are the primary tool used to detect and stage kidney tumors. However, traditional CT interpretation depends heavily on radiologists reviewing images manually, a process that is time-consuming, inconsistent between different doctors, and prone to human error. Small or subtle tumors can easily be missed, especially when radiologists face large volumes of scans.
A key challenge in kidney tumor diagnosis is not just detecting that a tumor is present, but also correctly classifying what kind of tumor it is. Some tumors are benign (non-cancerous) and may not require aggressive treatment, while others are malignant and may need surgery, targeted therapy, or immunotherapy. Knowing the specific subtype is critical for choosing the right treatment plan.
Artificial intelligence (AI) and deep learning have shown great promise in medical imaging by automatically detecting patterns in CT scans that humans might miss. This study introduces a new AI framework specifically designed to detect kidney tumors and classify them into subtypes with very high accuracy, potentially transforming how kidney cancer is diagnosed and staged.
The researchers developed a two-stage AI pipeline: a Feature Extraction Stage and a Training and Optimization Stage. First, CT scan images are processed by a specialized AI encoder that converts each image into a compact numerical representation capturing all key visual features of the tissue. These representations are then fed into machine learning classifiers that make diagnostic decisions.
The feature extraction is performed by the RAD-DINO-MAIRA-2 encoder, a powerful neural network based on the Vision Transformer (ViT) architecture. Originally trained on chest CT scans, the encoder's knowledge was transferred to kidney CT scans through a process called transfer learning. The encoder divides each image into small patches and processes them together to capture both local tissue details and broad structural patterns simultaneously.
The framework uses a four-level hierarchical classification structure that mirrors how a radiologist actually thinks: first determining whether a kidney is normal or has a tumor, then deciding if the tumor is benign or malignant, then classifying the specific benign subtype (such as adenoma or angiomyolipoma), and finally identifying the specific malignant subtype (such as RCC or secondary metastasis).
To identify the best classifier for each level, the researchers tested 32 different machine learning algorithms, including Support Vector Machines, k-Nearest Neighbors, Gaussian Process classifiers, and Multi-Layer Perceptrons (neural networks). Five different normalization techniques were also tested with each classifier to ensure the best combination was found at every level of the hierarchy.
The RAD-DINO-MAIRA-2 encoder uses a Vision Transformer (ViT) architecture, which processes images in a fundamentally different way from traditional convolutional neural networks. Rather than scanning the image with fixed filters, the ViT divides the CT scan into small image patches and learns how each patch relates to all other patches, capturing long-range relationships across the entire image.
A special learnable CLS (classification) token is added to the beginning of the patch sequence. As the image patches pass through multiple transformer layers, the CLS token aggregates information from all patches across the whole image. This global view allows the model to understand the broader context of what it is seeing, which is crucial for distinguishing subtle differences between benign and malignant tumors.
The encoder produces high-dimensional numerical representations called embeddings, which compactly capture the essential characteristics of each CT image, including texture, shape, and intensity patterns. These embeddings are then normalized using z-score standardization to ensure consistent scale across all images before being passed to the classification algorithms.
The dataset was split so that entire patient cases, not individual images, were separated into training and testing groups. This prevents data leakage, ensuring the model is always evaluated on patients it has never seen before. This approach produces a more realistic assessment of how the system would perform in a real hospital setting with new patients.
The framework was tested across 25 independent trials using the KAUH dataset of 8400 CT scan images from 120 patients. This repeated testing was designed to confirm that results were consistent and not due to chance. The overall framework achieved a maximum accuracy of 98.29% and a mean accuracy of 94.72% across all 25 trials, demonstrating both high peak performance and reliable consistency.
At the tumor type classification level (benign versus malignant), the Gaussian Process classifier achieved a perfect accuracy of 100% in its best trial and a mean accuracy of 95.82% across all 25 trials. This is the most critical diagnostic distinction, as it determines whether a patient needs cancer treatment, and the near-perfect performance at this level is particularly significant for clinical use.
At the malignant tumor subtype level, which distinguishes between RCC and secondary metastatic tumors, the Multi-Layer Perceptron (MLP) neural network achieved 100% accuracy in its best trial and a mean accuracy of 95.51%. Accurately identifying the malignant subtype is essential for determining appropriate treatment, as RCC and metastatic tumors require very different therapeutic approaches.
Even at the most challenging level, distinguishing between benign tumor subtypes such as adenoma, angiomyolipoma, and lipoma, the k-Nearest Neighbors classifier achieved 97.66% maximum accuracy and 94.67% mean accuracy. Collectively, these results demonstrate that the hierarchical framework performs exceptionally well at every diagnostic step, not just at the final classification stage.
The hierarchical design mirrors the actual clinical decision-making process that radiologists and oncologists follow. By first checking whether any tumor exists, then whether it is cancerous, and then identifying the specific subtype, the AI system provides a structured and transparent reasoning pathway. This makes the AI output much easier for clinicians to understand and trust compared to a single black-box prediction.
A key clinical benefit is that the system provides intermediate diagnostic outputs at each level of the hierarchy. For example, if the model determines a tumor is benign, the clinical team can immediately consider whether invasive biopsy or aggressive treatment is necessary. If the model identifies a malignant subtype, the patient can be fast-tracked to an oncology specialist. This step-by-step clarity enables faster, more targeted care decisions.
The hierarchical structure also handles the problem of class imbalance, which is a common challenge in medical datasets where some tumor types are much rarer than others. By isolating rare subtypes into later classification stages where they compete only against closely related classes rather than all categories at once, the model avoids being biased toward common diagnoses and performs more reliably on rare conditions.
The framework's high performance across 25 repeated trials demonstrates a level of reproducibility that is critical for clinical adoption. A system that works well in one trial but poorly in another is unsuitable for patient care. The narrow performance variability observed in this study signals that the framework is stable enough to be considered for integration into real-world radiology workflows.
This study demonstrates that an AI-driven hierarchical framework can detect and classify kidney tumors from CT scans with accuracy exceeding 98%, outperforming conventional single-stage deep learning approaches in both peak performance and consistency. By structuring the diagnostic process into four sequential levels, the framework achieves high accuracy at every step while remaining interpretable and clinically aligned.
The use of the RAD-DINO-MAIRA-2 encoder for feature extraction represents a meaningful advance. Transferring knowledge from chest CT training to kidney CT analysis demonstrates that state-of-the-art medical imaging encoders can generalize to new anatomical domains, expanding the potential reach of AI diagnostic tools beyond the organs they were originally designed for.
For patients, this technology could mean faster and more accurate diagnoses, reducing the time between imaging and treatment decisions. In settings where radiologist expertise is limited or imaging volumes are high, AI tools like this framework could help ensure that no early-stage tumor is missed and that each patient receives the most appropriate care based on a precise tumor classification.
Future work will focus on expanding the dataset to include more diverse patient populations and imaging equipment types, developing a user-friendly clinical interface, and conducting prospective trials to validate performance in real hospital settings. These steps are essential before the framework can be formally adopted in clinical practice, but this study represents a compelling proof of concept for AI-assisted kidney tumor diagnosis.