Kidney tumors are often discovered incidentally on CT scans done for other reasons. Early detection gives patients much better treatment options, but distinguishing benign (non-cancerous) tumors from malignant (cancerous) ones on imaging alone is difficult. Most kidney tumor research uses large publicly available datasets from Western countries, but patient populations and imaging equipment differ around the world.
This study created a new CT scan dataset from patients at King Abdullah University Hospital in Jordan - called the KAUH dataset - specifically to test deep learning models in a Middle Eastern population. Having datasets from diverse geographic regions is important to ensure AI tools work globally, not just in the populations where they were originally developed.
The KAUH dataset contains 8,400 CT images from 120 patients - 60 with kidney tumors and 60 without. Among the 60 tumor cases, 38 had benign tumors (including adenoma, angiomyolipoma, and lipoma) and 22 had malignant tumors (11 renal cell carcinoma and 11 metastatic tumors). Having both benign and malignant cases allows testing of how well the AI can make the clinically important distinction between the two.
The study tested several different deep learning architectures for two separate tasks: detection (does this image contain a tumor?) and classification (if a tumor is present, is it benign or malignant?). For detection, three models were compared - VGG16, ResNet50, and a custom CNN with 6 convolutional layers (CNN-6). For classification, a simpler custom CNN with 4 layers (CNN-4) was used.
For the tumor detection task, ResNet50 and CNN-6 each achieved 97% accuracy, substantially outperforming VGG16 which reached only 60%. The poor performance of VGG16 is notable and suggests that this architecture may not be well-suited to this type of medical imaging task, or that it requires more training data to perform at its best. Both ResNet50 and CNN-6 proved capable of reliably identifying whether a kidney tumor was present in a CT image.
ResNet50's strength comes from its use of 'residual connections' - pathways in the neural network that allow information to skip over layers. This design helps the model learn complex features without losing important information during training. CNN-6 is a simpler custom architecture but demonstrated competitive performance, suggesting that well-designed simpler models can match more complex off-the-shelf ones for specific tasks.
For classifying detected tumors as benign or malignant, the CNN-4 model achieved 92% accuracy. This level of performance is clinically significant because the distinction between benign and malignant kidney tumors directly determines whether a patient needs surgery or can be safely monitored. Unnecessary surgery carries real risks for patients, so accurate non-invasive classification matters greatly.
The 92% classification accuracy was achieved on a relatively small dataset by modern deep learning standards. Larger datasets and further testing across different patient populations would be needed before such a system could be used clinically. However, the result establishes that deep learning can learn to distinguish kidney tumor types from CT images with meaningful accuracy.
The primary contribution of this study is the creation and public release of the KAUH dataset, which adds regional diversity to kidney cancer AI research. The study demonstrates that deep learning models trained on data from Jordan can achieve high accuracy, supporting the feasibility of applying these tools in Middle Eastern healthcare settings.
Key limitations include the relatively small size of the dataset - 120 patients is modest for deep learning, and the imbalanced distribution between benign and malignant cases (38 vs 22) could affect how the model learns. The study also did not test whether models trained on Western datasets would perform equally well on the KAUH data, which would have directly tested the generalizability problem in reverse.
This work highlights that AI kidney tumor detection and classification is feasible in diverse geographic settings. Two architectures - ResNet50 and CNN-6 for detection, CNN-4 for classification - performed well on the new Jordanian dataset. The models achieved accuracy levels that, if reproducible in larger studies, could meaningfully assist radiologists in clinical practice.
Going forward, the field needs larger multi-site datasets from diverse populations, head-to-head testing of models trained in one region and applied to another, and prospective clinical trials to measure real-world impact on patient care. The KAUH dataset now provides a starting point for researchers who want to test and improve kidney cancer AI tools in Middle Eastern populations.