FracDetNet: Fracture Detection for Pediatric X-rays

• The paper introduces FracDetNet, a modified YOLOv8s that integrates dual-focus attention and multi-scale calibration for improved detection of subtle pediatric wrist fractures.
• Its architecture combines a CSPDarknet backbone, a PAFPN neck with DFA, and a multi-scale calibration head to refine feature extraction and scale-specific inference.
• Empirical evaluations on the GRAZPEDWRI-DX dataset demonstrate notable mAP gains and sharper fracture localization, confirming its clinical relevance.

Searching arXiv for FracDetNet and closely related papers to ground the article in current literature. FracDetNet is a fracture detection framework for pediatric wrist X-ray imaging that is formulated as an object detector and implemented as a modified YOLOv8s architecture. In the published usage of the name, it denotes a detector that integrates Dual-Focus Attention (DFA) in the neck and a Multi-scale Calibration (MC) module in the detection head in order to address subtle, morphologically diverse fractures under variable imaging conditions. On the GRAZPEDWRI-DX dataset, the reported system achieves a mAP$_{50-95}$ of 40.0\%, a mAP$_{50}$ of 63.9\%, and an improved fracture-specific mAP of 56.1\% (Sun et al., 27 Sep 2025).

1. Definition and problem setting

FracDetNet is a YOLOv8-based fracture detection network specifically redesigned for pediatric wrist X-ray imaging. The task is posed as object detection rather than segmentation: the input is a single 2D X-ray image of size $640\times640$, and the output is a set of fracture-region bounding boxes with class labels and confidence scores (Sun et al., 27 Sep 2025).

The framework is motivated by several difficulties that are explicit in the literature: fractures may be very subtle, highly variable in morphology, imaged under widely varying angles and patient postures, and captured with suboptimal image quality. The stated design objective is therefore to preserve sensitivity to fine local evidence such as thin fracture lines while also using broader anatomical context to reduce ambiguity from overlapping structures and projection changes (Sun et al., 27 Sep 2025).

The term has a potential naming ambiguity in the broader arXiv record. It is distinct from "FractalNet: Ultra-Deep Neural Networks without Residuals," which does not define any model named FracDetNet (Larsson et al., 2016). It is also unrelated to fractional-calculus-based segmentation or fractional differential equation solvers that appear in other contexts under superficially similar abbreviations or themes (Acharya et al., 2016, Firoozsalari et al., 2023, Antil et al., 2020). In the concrete published sense, FracDetNet refers to the medical X-ray detector introduced in 2025 (Sun et al., 27 Sep 2025).

2. Architectural composition

FracDetNet is structurally a modified YOLOv8s with three standard macro-components and two added modules. The baseline scaffold comprises a CSPDarknet backbone with SPPF, a PAFPN neck, and a detection head. The architectural novelty lies in inserting DFA into the neck and replacing the standard head refinement with the MC module (Sun et al., 27 Sep 2025).

The processing flow is described as follows. The input X-ray is resized to $640\times640$. The CSPDarknet backbone extracts multi-scale feature maps and ends with SPPF to increase receptive field via fast pyramid pooling. The resulting features are passed through the PAFPN neck, which propagates information bidirectionally across scales. Selected neck feature maps are then refined by DFA, which computes local and global self-attention in parallel and fuses them. The multi-resolution outputs are passed to the detection head, where each scale is first processed by MC before YOLOv8 prediction layers generate the final bounding boxes and class scores (Sun et al., 27 Sep 2025).

Component	Role	Placement
CSPDarknet + SPPF	Multi-scale feature extraction and receptive-field expansion	Backbone
Dual-Focus Attention (DFA)	Parallel local/global self-attention with fusion	Neck
PAFPN	Top-down and bottom-up feature aggregation	Neck
Multi-scale Calibration (MC)	Channel recalibration and multi-scale spatial refinement	Head

This organization suggests a deliberate separation of functions. DFA addresses representational enrichment before prediction, whereas MC recalibrates per-scale features immediately prior to box and class inference. A plausible implication is that FracDetNet treats contextual aggregation and scale-specific calibration as distinct optimization problems rather than collapsing them into a single attention block.

3. Dual-Focus Attention

DFA is designed to emphasize both local self-attention and global self-attention. The local branch operates directly on the incoming neck feature map $\mathbf{X}$, while the global branch works on a pooled lower-resolution representation. The global branch pre-processing is given as

$\mathbf{X}_{\text{global} = \mathrm{LayerNorm}\bigl( \mathrm{AvgPool}\bigl( \mathrm{GELU}(\mathrm{Linear}(\mathbf{X})) \bigr) \bigr).$

Queries are derived from the local branch, while keys and values are computed for both the local and global branches (Sun et al., 27 Sep 2025).

DFA incorporates learnable relative positional bias. For the local branch, the raw relative position between query and key locations is encoded as

$\mathbf{R}_{\text{local}(i,j) = \mathrm{Norm} \left( \left[ \frac{i - k}{H - 1}, \frac{j - l}{W - 1} \right] \right),$

and then transformed by an MLP: $\mathbf{b}_{\text{local}(i,j) = \mathrm{MLP}(\mathbf{R}_{\text{local}(i,j)) = \mathbf{W}_2 \cdot \bigl( \mathrm{ReLU}(\mathbf{W}_1 \cdot \mathbf{R}_{\text{local}(i,j) + \mathbf{b}_1 ) \bigr) + \mathbf{b}_2.$ Analogous definitions are used for the global branch (Sun et al., 27 Sep 2025).

Queries and keys are L2-normalized,

$\widehat{\mathbf{Q} = \frac{\mathbf{Q}{\|\mathbf{Q}\|_2}, \qquad \widehat{\mathbf{K} = \frac{\mathbf{K}{\|\mathbf{K}\|_2},$

and local and global similarities are then formed by dot products plus relative positional bias: $\mathbf{S}_{\text{local}(i,j) = \widehat{\mathbf{Q}(i,j) \cdot \widehat{\mathbf{K}_{\text{local}(i,j)^T + \mathbf{b}_{\text{local}(i,j),$

$_{50}$0

After Softmax normalization, the attended values are concatenated and projected: $_{50}$1 This residual form makes DFA an additive refinement block within the neck (Sun et al., 27 Sep 2025).

In the reported ablations, adding DFA alone to YOLOv8s improves mAP from 37.2\% to 38.9\% and mAP$_{50}$2 from 61.3\% to 63.8\% (Sun et al., 27 Sep 2025). The paper interprets this as evidence that jointly modeling local fine structure and global anatomical context improves discriminative power for subtle fracture detection. That interpretation is consistent with the detector’s clinical target domain, where isolated pixel-level discontinuities are often not semantically sufficient.

4. Multi-scale Calibration head

The MC module is inserted into the detection head and is composed of a Channel Block and a Multi-scale Block. Its stated purpose is to recalibrate features so that fracture-relevant channels and spatial patterns are emphasized before prediction (Sun et al., 27 Sep 2025).

The Channel Block computes global average-pooled and max-pooled descriptors: $_{50}$3

$_{50}$4

and applies them as channel attention: $_{50}$5 This is a channel-wise recalibration stage based on global descriptors (Sun et al., 27 Sep 2025).

The Multi-scale Block then applies depthwise convolutions with progressively larger effective receptive fields. It first computes

$_{50}$6

and then three branches: $_{50}$7 These are fused into a spatial attention process: $_{50}$8 The use of axial decomposition is explicit: large effective kernels are realized by $_{50}$9 and $640\times640$0 depthwise convolutions to reduce computational cost (Sun et al., 27 Sep 2025).

Ablation results attribute a stronger isolated gain to MC than to DFA: relative to the YOLOv8s baseline, MC alone yields 39.5\% mAP and 63.2\% mAP$640\times640$1, while the full DFA+MC configuration reaches 40.0\% mAP and 63.8\% mAP$640\times640$2 (Sun et al., 27 Sep 2025). This suggests that per-scale recalibration at the head is a major contributor to the final detector’s accuracy.

5. Dataset, training setup, and empirical results

FracDetNet is evaluated on the GRAZPEDWRI-DX dataset from the Medical University of Graz. The dataset contains 20,327 images from 6,091 patients over the period 2008–2018, with 74,459 image-level labels and 67,771 annotated targets in the form of bounding boxes covering various fracture types (Sun et al., 27 Sep 2025).

Training and implementation details are reported with moderate specificity. The framework is implemented in MMYOLO and MMDetection, using Python 3.8, CUDA 11.8, PyTorch 2.0.0, and Torchvision 0.15.1 on a single NVIDIA RTX 3090 GPU. The input resolution is $640\times640$3, the batch size is 16, the initial learning rate is 0.01, and initialization uses pre-trained weights from COCO (Sun et al., 27 Sep 2025).

Evaluation uses COCO-style mean Average Precision, including mAP$640\times640$4, mAP$640\times640$5, scale-specific mAP$640\times640$6, mAP$640\times640$7, mAP$640\times640$8, fracture-specific mAP, FLOPs, and parameter count (Sun et al., 27 Sep 2025).

Model	mAP$640\times640$9	mAP$640\times640$0	Params
YOLOv8s baseline	37.2%	61.3%	22.97 M
FracDetNet	40.0%	63.9%	24.96 M

The full reported FracDetNet results are mAP$640\times640$1 40.0\%, mAP$640\times640$2 63.9\%, mAP$640\times640$3 13.5\%, mAP$640\times640$4 35.5\%, mAP$640\times640$5 41.9\%, mAP$640\times640$6 56.1\%, FLOPs 0.016 T, and 24.96 M parameters (Sun et al., 27 Sep 2025). The corresponding YOLOv8s baseline reports are 37.2\%, 61.3\%, 10.4\%, 26.9\%, 40.6\%, 54.5\%, 0.014 T, and 22.97 M (Sun et al., 27 Sep 2025).

The improvement over baseline is described in the source as +7.5\% absolute in mAP$640\times640$7, +4.2\% absolute in mAP$640\times640$8, and +2.9\% absolute in fracture-specific mAP (Sun et al., 27 Sep 2025). The numerical difference from 37.2\% to 40.0\% is 2.8 percentage points, so the quoted “+7.5\%” should be read exactly as reported in the paper rather than reinterpreted here. A likely implication is that the source is using a relative-improvement convention for at least some summary statements, even though the table itself lists absolute scores.

The reported qualitative analysis uses Grad-CAM visualizations showing that the baseline sometimes diffuses attention over irrelevant bone structures, whereas FracDetNet focuses more sharply on fracture locations and yields tighter attention around the true fracture site (Sun et al., 27 Sep 2025). This does not constitute a formal localization metric, but it is presented as supportive evidence for the intended effect of DFA and MC.

6. Relation to adjacent research and nomenclature

FracDetNet belongs to a broader lineage of detection networks that modify standard backbones, necks, or heads to improve optimization or feature representation. In naming, however, it should not be conflated with several distinct lines of work.

First, it is not derived from "FractalNet: Ultra-Deep Neural Networks without Residuals" (Larsson et al., 2016). That paper develops self-similar ultra-deep CNNs with interacting subpaths, mean-based joins, and drop-path regularization, but it explicitly does not define any model called FracDetNet. The resemblance is nominal rather than architectural.

Second, it is unrelated to fractional-gradient image segmentation methods such as "On Image segmentation using Fractional Gradients-Learning Model Parameters using Approximate Marginal Inference" (Acharya et al., 2016), where the central object is fractional gradient estimation for edge detection and segmentation rather than a YOLO-style detector.

Third, it is unrelated to fractional differential equation solvers such as "deepFDEnet" (Firoozsalari et al., 2023), "Fractional Deep Neural Network via Constrained Optimization" (Antil et al., 2020), or "fTNN: a tensor neural network for fractional PDEs" (Ma et al., 25 Jun 2026). Those works use the term “fractional” in the sense of fractional calculus, Caputo derivatives, or fractional Laplacians, whereas FracDetNet uses “Frac” as shorthand for fracture detection (Sun et al., 27 Sep 2025).

There is, however, a thematic connection to domain-specific fracture analysis in other fields. For example, a convolutional network for prestack seismic fracture detection predicts fracture density, dip, and strike from azimuthal seismic responses (Yuan et al., 2021). That work addresses a different modality and target representation, but it illustrates that “fracture detection network” can denote structurally diverse systems across medical imaging and geoscience.

7. Significance, limitations, and interpretation

The reported significance of FracDetNet is that it improves fracture detection on a clinically relevant pediatric wrist X-ray benchmark while retaining computational characteristics consistent with practical deployment. Relative to YOLOv8s, FLOPs increase from 0.014 T to 0.016 T and parameters from 22.97 M to 24.96 M (Sun et al., 27 Sep 2025). The paper characterizes this overhead as modest in view of the reported accuracy gains.

The main limitations are also explicit. The method is trained and evaluated on a single institutional dataset, so domain shift across hospitals or imaging devices may require fine-tuning (Sun et al., 27 Sep 2025). The text also indicates that very subtle fractures, fractures obscured by overlapping bones, extremely poor image quality, and unusual projections remain challenging. Exact training schedule details and data augmentation policies are not fully enumerated in the provided description, which constrains reproducibility at the level of optimizer scheduling and augmentation ablations (Sun et al., 27 Sep 2025).

From a methodological perspective, FracDetNet’s contribution is modular rather than paradigmatic. It does not replace the YOLOv8 detection paradigm; instead, it augments it with one attention mechanism in the neck and one calibration mechanism in the head. This suggests that the framework is best interpreted as a task-specific architectural refinement for medical radiography rather than a general theory of fracture detection. The broader implication is that subtle-fracture detection benefits from a two-stage representational strategy: local/global contextualization before prediction, followed by scale-aware recalibration at the head.