Von Hippel-Lindau (VHL) syndrome is a hereditary genetic condition that dramatically increases a person's lifetime risk of developing kidney cancer, specifically clear cell renal cell carcinoma (ccRCC). People with VHL syndrome require lifelong imaging surveillance to monitor for new tumors and track existing ones. Because these patients need repeated imaging over decades, doctors prefer MRI over CT scans to avoid cumulative radiation exposure.
VHL kidneys are extraordinarily complex to analyze on imaging. Unlike sporadic (non-hereditary) kidney cancer - where a patient typically has one tumor - VHL patients often have dozens of tumors and cysts developing simultaneously in both kidneys, ranging widely in size and appearance. Distinguishing a small solid tumor from a simple cyst on MRI is challenging even for experienced radiologists, and manually measuring and tracking hundreds of lesions across years of scans is enormously time-consuming.
Automated AI tools to segment (outline and classify) tumors and cysts on MRI could dramatically reduce this burden. However, existing AI methods for kidney cancer segmentation were almost all developed for CT scans in sporadic RCC cases, not for the far more complex MRI appearances seen in hereditary VHL disease. This paper addresses that gap by developing and testing new AI segmentation methods specifically for VHL patients.
The research team developed a novel AI architecture called the Hinge Forest (HF), which combines the strengths of two different machine learning approaches. Like conventional deep learning, it uses convolutional neural network blocks (specifically VGG blocks) to extract image features automatically. However, instead of a standard neural network output layer, it uses a mathematically reformulated decision forest - a structure based on decision trees - that is fully differentiable and can be trained with standard backpropagation techniques.
The key innovation is a consistent differentiable logic system where positive real numbers represent "true" and negative numbers represent "false." This allows the traditional binary decisions in a decision tree (go left or go right based on a threshold) to be replaced with smooth, gradient-friendly operations using ReLU functions. The result is a model that thinks like a decision tree but learns like a neural network. The HF used in this study consisted of 100 depth-7 hinge trees working together as a forest.
This was compared against U-Net, the standard deep learning architecture for medical image segmentation. U-Net is a powerful but opaque "black box" - it makes accurate predictions but cannot explain why. HF offers an additional advantage: leaf maps, which visually display which parts of a decision tree were activated for each image region. These maps can reveal what image features the model is using to identify tumors versus cysts versus normal kidney tissue, providing a window into the model's reasoning that U-Net cannot offer.
The study used MRI data from 115 VHL patients (117 imaging studies) collected between 2015 and 2021 at the National Institutes of Health. This dataset contained an unprecedented 504 confirmed ccRCC tumors and 1,171 cystic lesions - far more complex than anything previously used in renal segmentation AI research. Images came from five different MRI scanner types, adding real-world variability to the test.
Each MRI scan included four contrast phases: precontrast, corticomedullary (20 seconds after contrast), nephrogenic (70 seconds), and excretory (3 minutes). All images were manually segmented by trained radiologists into three categories: kidney parenchyma (normal tissue), cysts, and tumors. These manual segmentations served as the ground truth for AI training and evaluation. Both HF and U-Net were trained and tested on 10 randomized splits (75% training, 25% testing) to ensure reliable performance estimates.
Performance was measured using Dice Similarity Coefficient (DSC), which measures how closely an AI segmentation overlaps with the expert manual segmentation. A DSC of 1.0 means perfect overlap; 0.0 means no overlap. Connected component analysis was also performed to count how many individual tumors were detected, missed (false negatives), and incorrectly marked (false positives) in each image.
For tumor segmentation, Hinge Forest achieved a mean DSC of 0.53 compared to U-Net's 0.46 - a statistically significant difference (p less than 0.01). This means HF correctly identified and outlined tumors substantially more accurately than the standard approach. In the best-performing test case, HF achieved a tumor DSC of 0.86 vs. 0.74 for U-Net. For cyst segmentation, HF also outperformed U-Net (DSC 0.44 vs. 0.41, also statistically significant).
However, U-Net performed significantly better at segmenting the kidney parenchyma as a whole-organ mask (DSC 0.78 vs. 0.75). This difference may be explained by architectural differences: U-Net's larger receptive field allows it to better capture the full extent of a large organ, while HF's smaller-receptive-field VGG blocks appear better suited to the localized, fine-grained features of small tumors. Both models struggled with cysts and tumors overall due to the complexity of VHL MRI data, motion artifacts, and the fact that many smaller cysts were not annotated in the training data.
As a tumor detector, HF achieved a higher AUC (0.68) than U-Net (0.57), though both models missed a substantial number of tumors. HF missed 397 of 773 total ground truth tumor components compared to U-Net's 552 missed out of 773. The leaf map analysis revealed that specific decision rules in HF could account for 76% of kidney parenchyma voxels and 39% of tumor voxels in the test set, demonstrating that the model had learned meaningful, anatomically coherent image features.
This study demonstrates that AI can meaningfully automate one of the most burdensome tasks in VHL surveillance: tracking and measuring dozens of tumors and cysts across repeated MRI scans. Even at current accuracy levels, automated segmentation tools could pre-populate measurements that radiologists then verify and correct, dramatically reducing the time required for each surveillance MRI review.
The interpretability of Hinge Forest through leaf maps is a significant advantage for clinical translation. In medicine, AI tools are more likely to be adopted when clinicians can understand and validate what the model is paying attention to. Leaf maps showing that specific decision rules consistently correspond to kidney parenchyma or tumor regions provide a basis for understanding and trusting the AI's output - something not possible with standard U-Net approaches.
The authors acknowledge key limitations: both models still miss a substantial fraction of tumors, performance is limited by the complexity and heterogeneity of VHL MRI data, and the study was conducted at a single institution. Future work should explore larger and more diverse datasets, better initialization strategies for hinge trees, and extensions toward predicting tumor growth rates - one of the most clinically important uses of automated segmentation in VHL management.