Invisible Yet Detected: PelFANet with Attention-Guided Anatomical Fusion for Pelvic Fracture Diagnosis

Technical Summary: PelFANet with Attention-Guided Anatomical Fusion for Pelvic Fracture Diagnosis

Problem Statement
Pelvic fractures are critical injuries often resulting from high-energy trauma, carrying significant risks of hemorrhage and mortality. Diagnosis relies heavily on radiographic evaluation, yet subtle or "invisible" fractures—those lacking obvious cortical disruptions on standard X-rays but confirmed via 3D-CT—are frequently missed, even by skilled radiologists. Existing deep learning approaches often struggle in these scenarios. While segmentation-guided classification helps focus on anatomical structures, it risks discarding crucial contextual cues (such as limb alignment, joint spacing, or soft tissue indicators) that lie outside the bone boundaries. Conversely, analyzing raw X-rays alone may fail to isolate specific anatomical details amidst background noise. The challenge lies in developing a model that effectively integrates localized anatomical precision with global contextual information to detect both overt and subtle fracture signs.

Methodology
The authors propose PelFANet (Pelvic Fused Attention Network), a dual-stream attention-guided architecture designed to fuse raw pelvic X-rays (PXR) with segmented bone images. The framework operates via a two-stage pipeline:

1. Bone Segmentation: A U-Net with a Mix Transformer B0 encoder is trained to generate binary bone masks from raw PXRs. This model treats the entire pelvic region as a single class, producing segmentation masks that serve as the second input stream.
2. Dual-Stream Classification (PelFANet):
   • Input: The network accepts two parallel inputs: the original raw PXR and the corresponding bone segmentation mask, both resized to 224×224.
   • Architecture: The core consists of two independent convolutional branches that extract features from each modality. These branches are interconnected through eight stacked Fused Attention Blocks (FABlocks).
   • Fusion Mechanism: Within each FABlock, feature maps from both streams are processed via stream-specific convolutions, concatenated, and refined using a Convolutional Block Attention Module (CBAM). The CBAM applies sequential channel and spatial attention to highlight informative features. The resulting fused features are projected and redistributed back into both streams via residual connections, enabling iterative cross-stream refinement.
   • Output: After the final FABlock, features undergo global pooling and are passed through a fully connected layer for binary fracture classification.

Key Contributions

• Dual-Input Architecture: The paper introduces a novel approach that simultaneously leverages raw image context and segmented anatomical structures, addressing the limitation of methods that rely solely on one modality.
• Attention-Guided Fusion: The design of FABlocks allows for dynamic interaction between the raw and segmented streams, enabling the model to learn feature combinations that capture both global context and localized anatomical details.
• Generalization to Invisible Fractures: The study specifically targets the detection of "invisible" fractures (subtle signs not visible on X-ray but confirmed by CT), a scenario where standard deep learning models often underperform.

Experimental Results
The model was evaluated on the AMERI dataset, which includes a Visible Fracture (VIS) subset and a challenging Invisible Fracture (INVIS) subset (fractures confirmed by CT but not visible on X-ray).

• Segmentation: The bone segmentation model achieved an average Intersection over Union (IoU) of 90.28% and a Dice Score of 92.78%.
• Classification Performance (VIS): On the visible fracture subset, PelFANet achieved 88.68% accuracy, 92.21% recall, and an AUC of 0.9334.
• Classification Performance (INVIS): Crucially, the model was tested on the INVIS subset without being trained on it. It achieved 82.29% accuracy, 84.35% recall, and an AUC of 0.8688.
• Comparison: PelFANet outperformed prior baselines (including ImageNet and DRR20 pretraining strategies) on both subsets. Notably, on the INVIS subset, it improved the AUC from 0.8002 (previous best) to 0.8688 and the F1 score from 78.60% to 81.23%.
• Visualization: Grad-CAM visualizations indicated that correctly classified cases showed focused activation on relevant bone regions, whereas misclassified samples exhibited scattered attention.

Significance and Claims
The authors claim that PelFANet demonstrates the clinical potential of anatomy-aware, dual-input architectures for robust fracture detection. The primary significance lies in the model's ability to generalize to invisible fracture cases despite being trained exclusively on visible fracture data. By retaining global structural context (from raw X-rays) alongside precise anatomical boundaries (from segmentation), the network captures subtle, non-local cues indicative of underlying injury that single-stream models miss.

The paper concludes that while the results are promising, further investigation is required to fully understand the specific contributions of components like CBAM and FABlocks through ablation studies. Additionally, the authors note that future work should involve validating the model on larger, multi-center cohorts and exploring modern architectures beyond the ResNet-based comparisons used in this study. The work aims to support robust real-time clinical use by addressing the diagnostic challenges of subtle radiographic presentations.