PelFANet: Dual-Stream Pelvic Fracture Diagnosis
- The paper introduces PelFANet, a dual-input architecture that fuses raw pelvic X-rays with segmented bone images via Fused Attention Blocks to enhance fracture detection.
- It employs a segmentation-guided, two-stage pipeline with CBAM to integrate global anatomical context and localized bone structure for subtle fracture identification.
- Empirical results demonstrate improved sensitivity, achieving an AUC of 0.9334 for visible fractures and 0.8688 for invisible fractures compared to prior methods.
Searching arXiv for PelFANet and closely related naming variants to ground the article in current papers. PelFANet is a dual-stream attention network for pelvic fracture diagnosis that fuses raw pelvic X-rays with segmented bone images in a segmentation-guided two-stage pipeline. It was introduced to address a specific radiographic failure mode: pelvic fractures that are subtle on standard radiographs or entirely invisible on X-ray but later confirmed on 3D-CT. The network employs Fused Attention Blocks (FABlocks) to iteratively exchange and refine features from both inputs, with the aim of preserving global anatomical context from the raw image while injecting localized anatomical structure from the segmented bone image (Bhuiyan et al., 17 Sep 2025).
1. Clinical setting and diagnostic target
Pelvic fractures are high-risk injuries commonly caused by high-energy trauma such as motor vehicle accidents or falls. They are clinically important because the pelvis protects major organs and blood vessels, so fractures can lead to hemorrhage, multi-organ damage, and substantial mortality. The reported in-hospital mortality rates are 5% to 20%, depending on injury severity and associated shock/injuries (Bhuiyan et al., 17 Sep 2025).
The central diagnostic problem is that standard pelvic radiographs are often insufficient in trauma settings. Even experienced radiologists may miss fractures, and up to 20% of pelvic fractures are initially overlooked. The paper distinguishes between visible fractures, where signs exist but may still be subtle or complex, and invisible fractures, where the X-ray shows no obvious fracture sign but a fracture is later confirmed on 3D-CT (Bhuiyan et al., 17 Sep 2025).
Several factors make invisible or subtle fractures difficult to detect on radiographs. Pelvic anatomy is complex, with overlapping bones and complex geometry that make cortical breaks hard to see. Radiographs have limited contrast and projection ambiguity, so a 2D X-ray can hide fracture lines or make them appear faint. The paper further emphasizes that subtle fractures may rely on non-local cues, including abnormal alignment, joint spacing, limb asymmetry, pubic symphysis widening, rotation, or dislocation. These cues may lie outside the segmented bone boundary, so a bone-only input can miss them. This motivates a model that reasons jointly over localized bone structure and broader anatomical context.
2. Dual-input formulation and FABlock mechanism
PelFANet is a dual-stream convolutional architecture with three main stages: parallel feature extraction, attention-guided fusion via stacked FABlocks, and final aggregation and classification. The network takes two inputs, a raw pelvic X-ray and the corresponding segmented bone image. Both are resized to 224×224 and processed in parallel (Bhuiyan et al., 17 Sep 2025).
Each stream begins with a 3×3 convolution, batch normalization, ReLU, and max pooling. This produces low-level features from the raw-image stream and the segmentation stream. The central mechanism is the Fused Attention Block, which performs interactive feature exchange between the two streams. Let the feature maps be
Each stream is first processed separately: The outputs are concatenated: $F_{cat} = \mathrm{Concat}(F_1, F_2) \tag{2}$ with
The concatenated representation is then refined using CBAM: where denotes the Combined Feature with Attention. The fused feature is redistributed back to both branches through residual updates:
This design gives the raw X-ray stream access to anatomical specificity from the segmented bone stream, while the segmentation stream acquires contextual awareness from the raw radiograph. The paper describes this as progressive cross-stream refinement. It explicitly states that the architecture includes eight FABlocks, with channel sizes shown as FABlock (64, 64), FABlock (64, 128), FABlock (128, 128), FABlock (128, 256), FABlock (256, 256), FABlock (256, 512), and FABlock (512, 512) (Bhuiyan et al., 17 Sep 2025).
After the final FABlock, the fused feature map passes through a 3×3 convolution, then CBAM, then batch normalization, then ReLU. Global features are extracted via adaptive average pooling and flattening into a 1024-dimensional vector, which is fed into a fully connected layer for binary classification: Fracture or Normal.
3. Segmentation-guided pipeline and anatomical supervision
PelFANet uses a segmentation-guided two-stage pipeline. In Stage 1, a segmentation network generates bone masks from raw pelvic radiographs. In Stage 2, the raw X-ray and the segmented bone image are jointly processed for fracture classification. The purpose of this decomposition is to inject anatomical guidance into classification while preserving whole-image context (Bhuiyan et al., 17 Sep 2025).
The bone segmentation module is a U-Net with a Mix Transformer B0 encoder. The full pelvic region is treated as a single class, and the model is trained as a binary segmentation task. The output masks are cropped to create the segmented bone image used as the second input to PelFANet. The segmentation training setup is specified as follows: ImageNet pretraining, input size 224×224, 2-fold cross-validation, Dice Loss, Adam, initial learning rate , cosine annealing, minimum learning rate , cycle length 50, 300 epochs, and batch size 25 (Bhuiyan et al., 17 Sep 2025).
The segmentation model achieved an Average IoU of 90.28% and an Average Dice Score of 92.78%. In the paper’s interpretation, these results support the reliability of the anatomical masks used by PelFANet. A plausible implication is that the segmentation branch is not merely a localization prior; it functions as a structured anatomical representation that conditions the downstream classifier on pelvic morphology rather than raw texture alone.
CBAM inside the FABlocks is composed of a Channel Attention Module and a Spatial Attention Module. Its stated role is to emphasize informative channels and relevant spatial regions in the fused representation. This makes the fusion mechanism more specific than simple concatenation, because the combined representation is filtered before being redistributed to both streams.
4. Training protocol and datasets
PelFANet is pretrained on the COVID-QU-Ex chest X-ray dataset to learn general radiographic features. The dataset contains 33,920 CXR images, partitioned into COVID-19: 11,956, Non-COVID infections: 11,263, and Normal: 10,701, with ground-truth lung masks available for all images. The pretraining split is 80% train and 20% test, with 20% of the training set used for validation. Each image is augmented 4× using rotation within 25°, shearing within 10%, horizontal flipping, and translation within 10%, producing an expanded training set size of 108,575 images. Pretraining uses CrossEntropyLoss, SGD, a learning rate of 0.0001, StepLR, a decay factor of 0.1 every 10 epochs, 100 epochs, and batch size 64 (Bhuiyan et al., 17 Sep 2025).
Fine-tuning is performed on the VIS subset of the AMERI dataset using reshuffled 5-fold cross-validation. The VIS subset contains 228 pelvic X-rays, with 168 fracture and 60 normal. To mitigate class imbalance, each fracture case was augmented into 2 variants and each normal case into 6 variants. During fine-tuning, the final classifier is changed from 3 outputs to 2 outputs, while the loss, optimizer, and scheduler remain the same as in pretraining; training uses 30 epochs and batch size 8 (Bhuiyan et al., 17 Sep 2025).
The AMERI dataset is the primary pelvic X-ray dataset for evaluation. Its visible fracture subset consists of 228 images collected from 315 subjects at Steel Memorial Hirohata Hospital in Japan, acquired between April 2013 and August 2019, with fractures confirmed by experienced radiologists and cases with implants or incomplete pelvic coverage excluded. The invisible fracture subset contains 23 fracture and 12 normal cases, where fractures are not visible on X-ray and are confirmed by corresponding 3D-CT. The reported evaluation metrics are Accuracy, Precision, Recall, Specificity, F1-score, and AUC.
5. Empirical performance and comparative results
PelFANet’s reported empirical results cover segmentation quality, classification on visible fractures, and generalization to invisible fractures. On the VIS subset, the model achieved Accuracy 88.68% ± 0.11%, Precision 92.49% ± 0.09%, Recall 92.21% ± 0.17%, Specificity 78.33% ± 1.27%, F1-score 84.71% ± 0.46%, and AUC 0.9334 ± 0.10. On the INVIS subset, despite being trained only on visible fractures, it achieved Accuracy 82.29% ± 0.01%, Precision 88.36% ± 0.07%, Recall 84.35% ± 0.07%, Specificity 78.33% ± 0.44%, F1-score 81.23% ± 0.06%, and AUC 0.8688 ± 0.04 (Bhuiyan et al., 17 Sep 2025).
| Evaluation set | AUC | F1-score |
|---|---|---|
| VIS | 0.9334 ± 0.10 | 84.71% ± 0.46% |
| INVIS | 0.8688 ± 0.04 | 81.23% ± 0.06% |
The paper compares PelFANet against prior ResNet-based methods using ImageNet, DRR20, ImageNet + DRR20, and ImageNet + DRR20_Full pretraining. On VIS, PelFANet slightly exceeds the best prior AUC, with 0.9334 versus 0.9290 for DRR20. On INVIS, the margin is larger: 0.8688 AUC versus 0.8002 AUC for DRR20, and 81.23% F1 versus 78.60% F1 (Bhuiyan et al., 17 Sep 2025).
The reported interpretation is that PelFANet has strong sensitivity to fractures but moderate specificity, meaning some normal cases are falsely flagged as fractures. The model’s performance on INVIS is presented as one of the paper’s strongest claims, because it suggests the network learns abstract contextual cues rather than only direct fracture edges. This suggests that the dual-input architecture is especially useful when fracture evidence is subtle or absent in the raw radiograph.
6. Novelty, limitations, and scope of the term
The paper characterizes PelFANet’s novelty through six elements: dual-input design, anatomy-aware fusion, the FABlock mechanism, explicit targeting of invisible fractures, segmentation-guided classification, and improved generalization to invisible fractures despite no training on them (Bhuiyan et al., 17 Sep 2025).
Its principal methodological distinction is that segmented anatomy is not used only for localization. Instead, the segmented bone image is treated as a separate classification input that participates in iterative attention-guided fusion. Relative to conventional single-stream classifiers, the intended balance is between global context from raw radiographs and localized structure from segmentation. A plausible implication is that PelFANet belongs to a broader class of anatomy-aware dual-input radiographic architectures rather than to ordinary late-fusion or mask-guided pipelines.
The paper also states several limitations and future directions. It does not provide a full ablation table; detailed ablations of CBAM and FABlocks are planned. Pretraining on chest X-rays was chosen for radiographic similarity and mask availability, but training from scratch should be studied. Comparisons were made using ResNet backbones to align with prior work, and future work should include more modern architectures. Further validation on larger, multi-center cohorts is needed, and the approach may be extended to other anatomical regions (Bhuiyan et al., 17 Sep 2025).
A common source of confusion is nomenclature. In the cited corpus, PelFANet denotes the pelvic fracture diagnosis model introduced in “Invisible Yet Detected: PelFANet with Attention-Guided Anatomical Fusion for Pelvic Fracture Diagnosis” (Bhuiyan et al., 17 Sep 2025). It should not be conflated with FEONet, the “Finite Element Operator Network” for parametric PDEs (Lee et al., 2023), or with the EPON scheduling framework DWBA-FL for federated learning over 50G-EPON (Ciceri et al., 2021). The similarity in naming is incidental; the underlying problem domains, architectures, and evaluation criteria are unrelated.