Abstract
Identifying rib fractures in 3D CT scans is a complex task due to the subtle and irregular nature of these injuries. Traditional spatial-domain models often suffer from spectral bias, overlooking the high-frequency details critical for fracture detection. This paper presents RF-DTR, a novel neural network that operates in the frequency domain using Discrete Cosine Transform (DCT) tokens. By utilizing a multi-stage refinement architecture and a specialized Mahalanobis-regularized loss, RF-DTR effectively captures fine-grained fracture characteristics. Experimental results on the RibFrac dataset demonstrate that RF-DTR achieves state-of-the-art accuracy and segmentation performance with high computational efficiency.
Introduction
The detection of rib fractures in 3D CT imaging is often described as a "needle in a haystack" problem. Fractures are typically minute, hollow, and irregular, making them difficult to distinguish from the complex thoracic background. Current AI approaches face two primary challenges: spectral bias and a lack of iterative refinement. Most voxel-based models prioritize low-frequency information (broad shapes) while ignoring the high-frequency signals (edges and fine details) that define a fracture. Additionally, unlike human clinicians who refine their observations through multiple passes, most AI models are "one-shot" predictors.
Analysis of feature distributions using UMAP reveals that CT volumes are significantly more discriminable in the frequency domain than in the spatial domain. Fractures that appear as noise in voxel space often stand out clearly in specific frequency bands. This insight forms the basis for the RF-DTR approach.
Methods
RF-DTR (Rib Fracture DCT Token Regression Network) addresses these issues through several innovations. First, it employs DCT Tokenization, converting 3D sub-volumes into frequency-domain tokens to directly analyze tissue characteristics. This allows the model to capture structural integrity more effectively than pixel-by-pixel methods, particularly for "hollow" fractures where the bone center appears empty.
The core architecture is an encoder-only Vision Transformer (ViT) with a Regressor-After-Gating (RAG) mechanism. At each of the four stages, a Gating module identifies potential fracture patches, and a Regressor predicts the corresponding DCT tokens for the fracture mask.
To ensure the model focuses on high-frequency details, a Mahalanobis-Regularized Regression Loss is implemented. This loss de-correlates frequency features and applies self-adaptive weights to difficult-to-learn components. Furthermore, a Unified Cross-Stage Penalty enforces hierarchical constraints, ensuring that deeper stages provide progressively more accurate and confident predictions, effectively replicating the clinician's refinement process.
Results
The model was evaluated on the RibFrac Challenge Dataset and a refined version, RibFrac-GTRefined, curated by the authors to reduce label noise using high-quality rib annotations from RibSeg v2.
RF-DTR significantly outperformed existing methods. On the refined dataset, it achieved an F1-score of 0.714, compared to 0.630 for the frequency-based PatchDCT and 0.526 for the spatial-based UNETR++. It also reached a Dice Coefficient of 0.787 for segmentation. Notably, RF-DTR is computationally efficient, requiring only 17.50 GFLOPs, which is lower than standard spatial models like UNETR (27.62B) or FracNet (94.87B).
Discussion and Conclusion
While the multi-stage design generally improves performance, the authors noted a "non-monotonic" trend when increasing the number of stages beyond three, suggesting a potential for overfitting or vanishing gradients in deeper configurations.
In conclusion, RF-DTR demonstrates that modeling rib fractures in the frequency domain provides a significant advantage over traditional spatial methods. By capturing high-frequency details and employing a multi-stage refinement strategy, the model achieves superior detection and segmentation performance while remaining computationally lightweight. Future research will focus on optimizing token selection and extending this frequency-domain approach to other bone pathologies.
Summary updated