Hip joint image quality screening based on the Diffusion Mamba model

Abstract

The scarcity of high-quality annotated datasets hinders the development of reliable AI for hip joint pathology diagnosis. To address this, we present Hip-35, a comprehensive 2D hip radiograph dataset for orthopedic abnormalities, with 35,000 curated synthetic images in 35 pathology categories and 76 expert-annotated clinical radiographic examinations. Our key innovation is a three-stage quality control framework ensuring anatomical fidelity and clinical relevance: (1) Chromatic artifact detection via RGB channel-difference thresholding with edge masking, significantly reducing contamination; (2) Multi-metric deblurring with Manhattan distance, cosine similarity, and SSIM-guided fusion; (3) Anatomical hashing deduplication with U-Net-derived ROI signatures and temporal checks. Leveraging Diffusion Mamba’s (DiM) efficiency, we generate and refine synthetic candidates through our quality pipeline, achieving 92% acceptance in radiologists’ blinded validation. The final dataset includes key clinical variations, such as femoral neck fractures, acetabular deformities, and post-surgical cases. By publicly releasing the dataset and our open-source quality toolkit, this work establishes a new benchmark for medical image augmentation, enabling reproducible AI diagnostic development for rare orthopedic conditions.

Background & Summary

Hip joint pathology diagnosis remains a critical challenge in orthopedic imaging, where accurate detection of fractures, dislocations, and degenerative changes directly impacts surgical planning and patient outcomes^1–5. While deep learning methods have demonstrated potential in automatic fracture detection^6–8, existing models suffer from insufficient generalization capabilities due to the scarcity of annotated medical imaging data. Current public datasets for hip joint analysis, such as the NIH DeepLesion dataset⁹, lack fine-grained pathological categorization and anatomical specificity, particularly for rare conditions like acetabular stress fractures. This data scarcity problem is exacerbated by the labor-intensive nature of expert annotations and privacy constraints in medical data sharing¹⁰.

Traditional data augmentation techniques, such as affine transformations¹¹ and intensity variations¹², fail to capture the complex morphological patterns of hip pathologies, such as the trabecular bone density gradients in occult fractures or the metallic artifact distributions in post-surgical implants. Recent advances in diffusion models¹³ are promising for medical image synthesis, yet their application to high-resolution radiographic image synthesis remains limited due to two fundamental challenges: 1) The quadratic computational complexity of transformer architectures¹⁴ becomes prohibitive for 1024 × 1024 resolution synthesis; 2) Existing methods struggle to preserve anatomical coherence in synthesized images¹⁵, particularly at bone-cartilage interfaces and fracture margins.

Building upon the recent success of Mamba-based architectures¹⁶ in sequential modeling, we adapt the Diffusion Mamba (DiM) framework¹⁷ for orthopedic imaging. DiM integrates state space models (SSMs)¹⁸ with diffusion models^19,20, achieving linear computational complexity while maintaining global anatomical consistency. Our work extends DiM’s capabilities through key innovations tailored for medical synthetic hip radiograph generation:

• Anatomical Constraint Injection: Structural priors are injected into the Diffusion Mamba framework through multi-directional scanning, learnable boundary-aware padding, and local feature enhancement, helping preserve joint morphology, cortical continuity, and pathology-relevant fine textures during radiographic image synthesis.

• Multi-Resolution Hierarchical Refinement: Mamba’s selective scanning enables our coarse-to-fine approach: low-resolution SSM blocks represent global pelvic anatomy (iliac alignment, acetabular position) while high-resolution streams refine local pathology details (fracture morphology, subchondral cysts).

These innovations can efficiently generate high-resolution synthetic X-ray images and preserve the biomechanical relationship among the femoral head, acetabulum, and supporting structures, which is important for preserving pathology-relevant radiographic structure. To ensure clinical application, we further introduced a quality assurance framework combining color difference detection, multi-metric deblurring, and perceptual hash deduplication, and verified it through quantitative indicators and clinical expert reviews.

We provide a more detailed overview of previous work, and below report existing methods in two key areas: the application of deep learning in hip joint pathology analysis and advanced methods for synthetic data generation.

Deep learning in hip joint pathology analysis

Deep learning approaches have shown significant promise in automatic detection and classification of hip joint pathologies. Mutasa et al.²¹ proposed a CNN-based approach for hip fracture detection, achieving high accuracy on standard datasets. However, these models are often constrained by rare pathologies or subtle morphological changes due to limited training data. Public datasets like NIH DeepLesion⁹, while extensive, lack the fine-grained annotations required for nuanced hip joint analysis, such as differentiating early-stage osteoarthritis and normal variants²². To alleviate the data scarcity problem, recent studies^23–25 have explored transfer learning for general medical imaging datasets, but domain-specific challenges remain, including unique anatomy²⁶ and artifacts²⁷, which limit generalization.

Synthetic data generation based on diffusion model

Traditional data augmentation methods fail to replicate the complex morphological patterns of hip pathologies. For instance, they cannot accurately simulate trabecular bone patterns in stress fractures or metallic artifacts in post-surgical implants. Generative models, particularly GANs²⁸ and diffusion models, offer more sophisticated augmentation capabilities. GANs like CycleGAN²⁹ have been used for image-to-image translation, but they often produce artifacts and lack anatomical fidelity, especially at tissue boundaries³⁰.

Diffusion models¹³ have been applied to high-quality medical images generation. However, their application to high-resolution X-ray imaging is limited by computational constraints^31,32. Transformer-based diffusion models¹⁴, while powerful, suffer from quadratic complexity, making them impractical for large-scale medical image synthesis. Furthermore, maintaining anatomical consistency, especially at key interfaces such as bone-cartilage junctions, remains a challenge for existing diffusion frameworks³³. State Space Models (SSMs) provide a solution to these challenges. The Mamba architecture¹⁶ achieves linear computational complexity through a selective scanning mechanism while maintaining the global receptive domain. The combined architecture of the diffusion framework¹⁷ and the Mamba architecture decomposes 2D images into multi-directional 1D sequences by altering scanning modes (horizontal, vertical, diagonal), and extends these advantages to image generation. Our method builds on the latest advancements in hierarchical diffusion¹⁷ to limit multi-scale anatomical constraints. By combining SSMs with diffusion models, global anatomical structures and fine-grained pathological details are captured, addressing the limitations of existing methods in preserving tissue interfaces and fracture margins. This innovation associated with rare hip lesions can address the problem of limited real-world hip joint images and improve the generalizability of deep learning models.

Methods

This study introduces a hybrid real-synthetic 2D hip radiograph dataset designed to advance the development of AI-driven diagnostic models for hip joint pathologies. The dataset is constructed using a two-stage strategy: (1) curation and annotation of real hip radiographic examinations, and (2) large-scale synthetic data augmentation via the Diffusion Mamba (DiM) model–a novel diffusion framework built upon SSMs. The complete pipeline for generating and curating the Hip-35 dataset is depicted in Fig. 1.

Fig. 1.

Overview of the Hip-35 dataset generation and quality-control pipeline. The pipeline starts from 76 clinical radiographic cases covering 35 pathology categories with expert annotations. A Diffusion Mamba framework is first pre-trained on 256 × 256 patches and then fine-tuned on 512 × 512 patches to generate higher-resolution synthetic X-ray images. The generated images are subsequently processed by a quality-inspection module, including color-artifact detection, blur filtering, and deduplication, to separate accepted images from low-quality images. The upper panels illustrate the main architectural components of the Diffusion Mamba model, including scan patterns, padding strategy, and tokenization.

Original Dataset

The source clinical dataset consisted of 76 de-identified hip radiographic examinations retrospectively collected from the Department of Orthopedics, Shuozhou People’s Hospital, in a single-center retrospective cohort, covering 35 distinct clinical pathology categories, including normal hip joints, dislocations, stress fractures, femoral neck fractures, acetabular deformities, and post-surgical conditions. Representative examples are shown in Fig. 2. The source radiographic examinations were acquired using digital radiography (DR) systems from Beijing Wandong and Philips. These source clinical images were used for expert annotation and subsequent synthetic image generation. All source images were reviewed and annotated under a double-blind protocol by three board-certified radiologists, each with over five years of experience in musculoskeletal imaging, to ensure label consistency and clinical relevance. The original 76 clinical radiographic examinations are not part of the public data release and are not publicly redistributed through this study. The public release associated with this work is limited to the Hip-35 dataset described below.

Fig. 2.

Representative hip radiographic examples with various conditions.

Synthetic Image Generation via DiM

In the synthetic data generation phase, we employed the cutting-edge Diffusion Mamba (DiM) architecture¹⁷. DiM overcomes the computational bottlenecks faced by traditional transformer-based models when processing high-resolution images. The computational complexity of traditional models scales quadratically, making the generation of high-resolution medical images expensive and time-consuming. In contrast, DiM, with its innovative SSM core, achieves linear computational complexity, showing an unparalleled advantage in efficiently generating high-fidelity, high-resolution images. We adapted the core DiM framework in this paper to better model the global anatomy and fine pathological details of hip radiographs.

The Mamba model was originally designed to process one-dimensional sequential data. This is achieved through its foundation in SSMs, which map a continuous input signal x(i) to a hidden state h(i) and then to an output y(i) using ordinary differential equations (ODEs):

$$h^{\prime }\left(i\right)=\mathbf{A}h\left(i\right)+\mathbf{B}x\left(i\right),y\left(i\right)=\mathbf{C}h\left(i\right)+\mathbf{D}x\left(i\right),$$ 1

where A, B, C, D are the weights of the SSM. For discrete data like image patches, these ODEs are discretized using rules like the zero-order hold (ZOH), allowing the model to be solved recurrently:

$$h_i=\tilde{\mathbf{A}}{h}_{i-1}+\tilde{\mathbf{B}}{x}_i,y_i=\mathbf{C}{h}_i+\mathbf{D}{x}_i.$$ 2

To enable the understanding and generation of 2D hip X-ray images with complex spatial relationships, we adapted several key architectural innovations of DiM.

Multi-Directional Scanning

To overcome the “causal bias” of a simple raster scan, which restricts the receptive field, we implement a multi-directional scanning mechanism. Let the 2D grid of patch features be denoted as $P\in {\mathbb{R}}^{H\times W\times D}$, where H and W are the number of patches vertically and horizontally, and D is the feature dimension. This mechanism defines a set of four distinct mapping functions ${\left\{{\mathcal{S}}_k\right\}}_{k=1}^4$. Each function ${\mathcal{S}}_k:{\mathbb{R}}^{H\times W\times D}\to {\mathbb{R}}^{\left(H\cdot W\right)\times D}$ flattens the 2D grid P into a unique 1D sequence X_k. These sequences correspond to row-major, reverse row-major, column-major, and reverse column-major orderings. By altering these scanning patterns across different Mamba blocks, the model ensures that every patch can causally attend to all other patches, thus achieving a true global receptive field critical for maintaining long-range anatomical consistency.

Learnable Padding Tokens

When flattening a 2D patch grid into a 1D sequence, spatial continuity is broken at the boundary of each row or column. To address this, we introduce a learnable padding token, $p_{pad}\in {\mathbb{R}}^D$, which is a trainable parameter vector. For a given scan direction, let a sub-sequence corresponding to one row or column be X_sub = (p₁, p₂, …, p_L). This sequence is augmented by appending the padding token into $X_{sub}^{\prime }=\left(p_1,\dots ,p_L,p_{pad}\right)$. These augmented sub-sequences are then concatenated. This mechanism functions as an explicit “end-of-line” or “end-of-column” signal, allowing the model to learn and recognize the 2D boundaries of the image, thereby preventing it from incorrectly associating the end of a femoral head in one row with the beginning of a pelvic organ in the next, effectively preventing spatial distortions and anatomical misalignments.

Local Feature Enhancement

The patchifying and flattening operations, while necessary for the Mamba architecture, can disrupt the immediate local neighborhood information of the original image. To compensate for this and preserve fine-grained details vital for diagnosis (e.g., sharp cortical bone edges, fine trabecular textures), we employ a lightweight local feature enhancement module, f_enh. This module consists of a standard 3 × 3 depth-wise convolution operation (DWConv_3×3). Let the initial patch features after the embedding layer be P_in. Before being processed by the Mamba blocks, these features are refined as:

$$P_{in}^{\prime }=f_{enh}\left(P_{in}\right)={\mathrm{DWConv}}_{3\times 3}\left(P_{in}\right).$$ 3

A symmetric operation is applied to the output features after the final Mamba block. This design re-injects critical local spatial information with minimal computational overhead, significantly enhancing the clarity and detail fidelity of the generated images.

Hierarchical Training and Generation Strategy

To maximize both training efficiency and final image quality, we designed and implemented a “Weak-to-Strong” hierarchical training and generation strategy.

Low-Resolution Pre-training

We first pre-trained the model on radiographic images that were downsampled to a 256 × 256 resolution. Training at a lower resolution is computationally inexpensive, allowing the model to quickly and efficiently learn the fundamental anatomical shapes, the general layout of the hip joint, and the coarse features of various pathologies, avoiding premature attention to high-frequency noise and minute details.

High-Resolution Fine-tuning

Once the model had mastered a solid structural foundation through pre-training, we moved it to fine-tuning on the native 512 × 512 resolution images. At this stage, the model could focus its full capacity on learning and refining high-resolution textures, complex pathological features (like occult fracture lines), and the natural transitions between different tissue types. This enables much faster convergence and a higher final image quality than training from scratch at high resolution.

Training-Free Upsampling

To generate even higher, 1024 × 1024 ultra-high-resolution images, we utilized an advanced, inference-time technique for upsampling that requires no additional training. This method allows us to generate images at resolutions the model has never been trained on, effectively overcoming the limitations of GPU memory and high-resolution training data. The principle is to use guidance in the early stages of the diffusion process to establish a coherent global anatomical framework. This is achieved using the upsample-guidance mechanism, formally expressed as:

$${\epsilon }_{\theta }\left(x_t,t\right)={\epsilon }_{\theta }\left(x_t,t\right)+{\omega }_t\mathbf{U}\left[\frac{1}{m}{\epsilon }_{\theta }\left(\frac{1}{\sqrt{P_t}}\mathbf{D}\left[x_t\right],\tau \right)-\mathbf{D}\left[{\epsilon }_{\theta }\left(x_t,t\right)\right]\right],$$ 4

where ε_θ is our noise-prediction model, m is the upscaling factor, U is the upsampling operator, D is the down-sampling operator, and ω_t is the guidance weight. In the later diffusion time steps, the model directly predicts noise on the high-resolution input to generate a synthetic radiographic image with pathological detail features.

Post-Generation Quality Control

To ensure the clinical relevance and anatomical fidelity of the generated images, each candidate image undergoes processing through a rigorous, multi-stage quality control pipeline. This includes chromatic artifact detection, multi-scale deblurring, and perceptual hashing for deduplication, which is formally outlined in Algorithm 1. Details of each stage are described below.

Algorithm 1

Integrated Quality Control Pipeline.

Chromatic Artifact Detection

In the chromaticity artifact detection module, the input three-channel image $I\in {\mathbb{R}}^{H\times W\times 3}$ is first processed through color space standardization:

$$I_{rgb}=\mathrm{BGR2RGB}\left(I\right).$$ 5

Subsequently, an edge mask is constructed to eliminate the interference of the peripheral air region, and its mathematical definition is:

$$M_e\left(x,y\right)={\mathbf{1}}_{\left[20,W-20\right]\times \left[20,H-20\right]}\left(x,y\right).$$ 6

This mask effectively avoids common annular artifacts by limiting the detection area to the pixel range of [20, W − 20] × [20, H − 20] from the image center. Based on the standardized RGB images, we calculate the maximum chromaticity difference between channels pixel by pixel:

$$D\left(x,y\right)=\underset{c\in \left\{R,G,B\right\}}{\max }{I}_{rgb}\left(x,y,c\right)-\underset{c\in \left\{R,G,B\right\}}{\min }{I}_{rgb}\left(x,y,c\right),$$ 7

and generate the pollution mask through the joint threshold determination:

$$M_c\left(x,y\right)={\mathbf{1}}_{D\left(x,y\right)>{\tau }_c}\cdot M_e\left(x,y\right).$$ 8

The final pollution decision function is defined as:

$$\mathrm{Contaminated}=\left[\frac{1}{HW}\sum _{x,y}{M}_c\left(x,y\right)\right]>{\tau }_p.$$ 9

Specifically, the local threshold is set to τ_c = 12 and the global area threshold is set to τ_p = 0.0005, which can detect at least 0.05% of significant abnormal areas.

Multi-Metric Feature Fusion Deblurring

In the hierarchical feature extraction architecture, we employ a ResNet-101 model to fuse multi-scale features, constructing a dual-path feature fusion mechanism as follows:

$$\mathrm{ℱ}\left(x\right)=\underset{\mathrm{Channel}\mathrm{Attention}\mathrm{Enhancement}}{\underbrace{\mathrm{SE\; -\; Block}\left(\mathrm{Conv4}\left(x\right)\right)}}\oplus \underset{\mathrm{Global}\mathrm{Semantic}\mathrm{Modeling}}{\underbrace{\mathrm{Conv5}\left(x\right)}}.$$ 10

The channel attention weights are dynamically generated through a squeeze-and-excitation mechanism, with the specific calculation process expressed as:

$$w_c=\sigma \left({\mathbf{W}}_2\cdot \mathrm{ReLU}\left({\mathbf{W}}_1\cdot \mathrm{GAP}\left(z_c\right)\right)\right).$$ 11

This design captures channel-level statistical features through global average pooling (GAP), then generates channel weights via dimensionality reduction → activation → dimensionality increase operations, ultimately achieving dynamic calibration of feature maps. In terms of defocus blur metrics, we propose a hybrid distance function to comprehensively evaluate image quality:

$$d_{\mathrm{hybrid}}=\alpha \underset{\mathrm{Intensity}\mathrm{Difference}}{\underbrace{{\parallel I_1-I_2\parallel }_1}}+\beta \underset{\mathrm{Structural}\mathrm{Dissimilarity}}{\underbrace{\left(1-\frac{{\mathbf{v}}_1^{\top }{\mathbf{v}}_2}{\parallel {\mathbf{v}}_1\parallel \parallel {\mathbf{v}}_2\parallel }\right)}}+\gamma \underset{\mathrm{Perceptual}\mathrm{Inconsistency}}{\underbrace{\left(1-\mathrm{SSIM}\left(I_1,I_2\right)\right)}},$$ 12

where the Manhattan distance measures pixel-level intensity differences, the cosine distance evaluates structural deviations in the feature space, and the SSIM term captures the perceptual characteristics of the human visual system. The optimal weight coefficients are determined through a constrained optimization process:

$$\begin{array}{c}\hfill \underset{\alpha ,\beta ,\gamma }{\min }\sum _{i=1}^N{\parallel f_{\theta }\left(I_i\right)-y_i\parallel }_2^2+\lambda \left({\alpha }^2+{\beta }^2+{\gamma }^2\right),\hfill \\ \hfill \mathrm{s.t.}\alpha +\beta +\gamma =1,\alpha ,\beta ,\gamma \ge 0.\hfill \end{array}$$ 13

These constraints ensure a balanced contribution from each metric term. Grid search validation shows that when λ = 0.1, the model achieves optimal balance on the validation set (α = 0.38, β = 0.29, γ = 0.33). Notably, the SSIM term uses an 11 × 11 Gaussian window with a standard deviation of 1.5 to match the local texture characteristics of medical images.

Perceptual Hashing Deduplication

In the image deduplication module, we constructed a multi-scale hash verification system to eliminate duplicate or highly similar medical image data. Global aware hashing achieves rapid initial screening through frequency-domain feature compression of the discrete cosine transform:

$$H_g=\mathrm{sign}\left({\mathrm{DCT}}_{64}\left(I_{\downarrow }\right)\right)\in {\left\{0,1\right\}}^{64},$$ 14

where the input image I undergoes double or triple downsampling to a resolution of 32 × 32, then its DCT low-frequency component is extracted and binarized into a 64-bit hash code. Local difference hashing focuses on the gradient structure characteristics of the 9 block regions:

$$H_l^k=\mathrm{rank}\left({\nabla }_x{I}_k\odot {\nabla }_y{I}_k\right),k=1,\dots ,9.$$ 15

We then calculate the Hadamar product -of the horizontal and vertical gradient maps within each 128 × 128 block, and generate an 8-bit hash fragment through sorting encoding. Anatomical ROI hashes further integrate the deep semantic features of the U-Net encoder:

$$H_r={\mathrm{AvgPool}}_{8\times 8}\left({\mathrm{U}‐\mathrm{Net}}_{\mathrm{enc}}\left(I\right)\right)\in {\mathbb{R}}^{256}.$$ 16

This hash value captures the potential representations of key anatomical patterns, such as the morphology of the femoral head and the distribution of trabeculae. The spatio-temporal consistency decision model integrates the above characteristics and temporal context information, and defines the retention probability as:

$$P_{\mathrm{keep}}=\frac{1}{Z}\exp \left(-\frac{\mathrm{\Delta }t}{24}\right)\cdot \exp \left(-0.7\parallel H_r^{\left(1\right)}-H_r^{\left(2\right)}{\parallel }_2^2\right),$$ 17

where the constant 24 in the temporal exponential term was heuristically chosen as a same-day decay scale, so that image pairs with short timestamp intervals receive a substantially stronger temporal penalty. The coefficient 0.7 in the anatomical similarity term was heuristically chosen to control the sensitivity to the ROI-feature distance. Smaller values reduce sensitivity to anatomical similarity and may retain near-duplicate images, whereas larger values may over-penalize subtle but meaningful anatomical variations. Therefore, 0.7 was adopted as a practical balance between duplicate suppression and preservation of anatomical diversity.

Overlap-Based Screening Evaluation

To assess the agreement between automatic post-generation quality control and manual review, we constructed a reference set for each pathology category by manually curating high-quality images under expert supervision. This reference set was then compared with the image set retained by the automatic quality-control pipeline. For each pathology category, we report the intersection size, union size, reference-set size, and selected-set size. We further define an overlap-based accuracy index as:

$$\mathrm{Accuracy\; Index}=\frac{\mid R\cap A\mid }{\mid R\cup A\mid }$$ 18

where R denotes the manually curated reference set and A denotes the automatically selected image set. This evaluation measures agreement between automatic screening and manual curation.

Final Dataset Composition and Quality-Control Screening Evaluation

Starting from the 76 expert-annotated source examinations, the DiM generation pipeline produced 1,000 synthetic images for each of the 35 pathology categories, resulting in a dataset of 35,000 generated images. Representative examples are shown in Fig. 3, which illustrates the visual diversity of the generated samples across different pathological patterns. All generated images were subsequently processed by the three-stage quality-control pipeline described above to remove chromatic artifacts, reduce excessive blur, and filter near-duplicate cases.

Fig. 3.

Representative examples of synthetic images generated by the DiM pipeline across multiple hip pathology categories.

To evaluate the agreement between automatic quality control and supervised manual curation, we compared the automatically selected image set with a manually curated reference set established under expert supervision for each pathology category. The corresponding overlap statistics are summarized in Table 1, including the intersection size, union size, reference-set size, result-set size, and the overlap-based accuracy index. Here, Result Size denotes the number of images retained by the automatic post-generation quality-control pipeline for a given pathology category, whereas Reference Size denotes the size of the manually curated reference set established under expert supervision.

Table 1.

Overlap-based evaluation of quality-control screening across all pathology categories.

Pathology Category	Accuracy Index	Intersection	Union	Reference Size	Result Size
After internal fixation of right femur	0.873	359	411	380	390
Bilateral hip joint deformity	0.837	309	369	334	344
Bone hyperplasia in right acetabulum	0.830	372	448	391	429
Comminuted fracture of right femoral intertrochanteric	0.844	358	424	374	408
Consider left inferior pubic ramus fracture	0.763	360	472	373	459
Fracture of left ascending ramus of pubic bone	0.827	350	423	363	410
Fracture of right ascending pubic ramus	0.793	357	450	369	438
Fracture of left ascending/descending pubic bone	0.856	375	438	399	414
Fracture upper right femur, normal left hip	0.847	387	457	404	440
Invisible sacral fissure	0.849	349	411	372	388
Left femoral head necrosis	0.834	352	422	373	401
Left femoral intertrochanteric fracture	0.819	339	414	364	389
Left femoral intertrochanteric fracture w/density	0.841	349	415	374	390
Left femoral intertrochanteric fracture w/tear	0.881	378	429	392	415
Local cortical discontinuity right iliac spine	0.849	393	463	423	433
No abnormality in pelvis	0.830	376	453	400	429
No obvious abnormalities in pelvis	0.779	318	408	338	388
Poor cortical continuity right pubic ramus	0.768	294	383	307	370
Postop right calcaneus & femoral neck fracture	0.821	343	418	366	395
Postop internal fixation left femur	0.862	350	406	372	384
Postop right femoral intertrochanteric fracture	0.893	411	460	429	442
Postop internal fixation right femoral neck	0.823	340	413	368	385
Postop internal fixation right femur	0.816	341	418	351	408
Postop left hip joint replacement	0.822	332	404	346	390
Removal internal fixation left femur	0.816	360	441	379	422
Right femoral fracture after left fixation	0.855	355	415	373	397
Right femoral neck base fracture	0.841	390	464	423	431
Right femoral neck fracture	0.872	346	397	368	375
Right ilium fracture	0.845	382	452	410	424
Linear shadows left iliac/pelvic area	0.840	367	437	403	401
Suspected right acetabular fracture	0.752	454	604	599	459
Left femoral neck fracture	0.829	358	432	375	415
Left hip joint dislocation	0.834	381	457	409	429
Left pubic bone fracture	0.861	427	496	448	475
Post right hip joint replacement	0.760	351	462	365	448
Overall Average	0.830	361.8	436.2	386.1	411.9
Standard Deviation	0.034	30.4	39.1	46.0	27.9

Across all 35 pathology categories, the overlap-based accuracy index remained consistently high, with an overall mean of 0.830, indicating substantial agreement between the automatic quality-control pipeline and the supervised reference curation. The highest agreement was observed for postop right femoral intertrochanteric fracture (0.893), followed by left femoral intertrochanteric fracture with tear (0.881) and after internal fixation of right femur (0.873). In contrast, suspected right acetabular fracture (0.752), post right hip joint replacement (0.760), and consider left inferior pubic ramus fracture (0.763) showed relatively lower agreement. These lower-scoring categories likely reflect greater structural variability, subtler pathological cues, or stronger postoperative heterogeneity, which make consistent filtering more challenging.

In summary, Fig. 3 and Table 1 show that the proposed pipeline can generate visually diverse pathological images while preserving stable agreement with supervised manual curation after post-generation quality control.

Data Record

The Hip-35 dataset is available at the Hugging Face Hub³⁴. It is organized into 35 folders, each corresponding to a specific pathology category. A complete list of all 35 categories is provided in Table 1, while the dataset’s file structure is illustrated in Fig. 4. Each folder contains 1,000 synthetic radiographic images in PNG format. The source clinical radiographic examinations used to generate this dataset are not included in the public release.

Fig. 4.

An example of the Hip-35 dataset’s file structure. The diagram illustrates the hierarchical directory layout, showing representative category folders and the naming convention for the 1,000 images contained within each.

Technical Validation

Experimental Settings

The experimental framework is implemented on an NVIDIA RTX 4090 GPU with 32GB memory, utilizing PyTorch 2.0 for medical image processing. The original dataset comprises 76 hip radiographic examinations, anonymized and annotated by three board-certified radiologists across 35 pathology categories, including femoral neck fractures, acetabular dysplasia, and post-surgical revisions. Synthetic images are generated at 1024 × 1024 resolution using the proposed DiM variant, with training conducted in two phases: initial pretraining on downsampled 256 × 256 images for 50 epochs (batch size 32, learning rate 3e-4), followed by gradient-aware fine-tuning at native 512 × 512 resolution for 30 epochs (batch size 16, learning rate 1e-4). AdamW optimization with cosine annealing schedules the learning rate decay.

Representative Overlap-Based Screening Results

To provide representative examples of the post-generation screening results, Table 2 summarizes overlap-based quality-control screening statistics for several selected pathology categories. These results are reported using the same overlap-based accuracy index defined above, which measures agreement between the automatically retained image set and the manually curated reference set. The “Total” column denotes the total number of generated synthetic images per pathology category before post-generation quality control (1,000 for each category).

Table 2.

Representative overlap-based screening results across selected pathology categories.

Pathology Category	Accuracy Index	Intersection	Union	Reference Size	Result Size	Total
Post right hip joint replacement	0.760	351	462	365	448	1000
Bilateral hip joint deformity	0.837	309	369	334	344	1000
Right femoral neck fracture	0.872	346	397	368	375	1000
Left femoral intertrochanteric fracture	0.819	339	414	364	389	1000
Suspected right acetabular fracture	0.752	454	604	599	459	1000
Left femoral head necrosis	0.834	352	422	373	401	1000
Based on all 35 pathology categories
Overall Average	0.830

Across the representative categories shown in Table 2, the overlap-based accuracy index ranged from 0.752 to 0.872. Higher agreement was observed for categories such as right femoral neck fracture (0.872), whereas lower agreement was found for suspected right acetabular fracture (0.752), likely reflecting subtler pathological cues and greater structural variability. The overall mean overlap-based accuracy index across all 35 pathology categories was 0.830, indicating substantial agreement between the automatic quality-control pipeline and supervised manual curation. Complete category-level statistics are provided in Table 1.

Quality Control Impact Analysis

Our quality control pipeline is designed to identify and filter out images with common artifacts, such as those shown in Fig. 5. The contribution of each quality control module was quantified through systematic ablation. Table 3 shows the progressive improvement from baseline to full system implementation.

Fig. 5.

Types of images that need to be cleaned.

Table 3.

Quality control module ablation study.

Configuration	Accuracy Index	Expert Acceptance	Time (ms)
Baseline (No QC)	0.538	0.67	142
+ Chromatic Artifact Detection	0.562	0.74	157
+ Multi-metric Deblurring	0.581	0.82	189
+ Perceptual Hashing Deduplication	0.595	0.88	213
Improvement	+0.057	+0.21	+71

Chromatic artifact detection significantly improved diagnostic accuracy while reducing expert rejection rates. Multi-metric deblurring substantially enhanced detection performance for subtle fracture patterns. Perceptual hashing deduplication effectively eliminated redundant cases without compromising quality. The complete quality control system demonstrated noticeable accuracy improvements with moderate computational overhead. Chromatic artifact detection improved diagnostic accuracy by 0.024 and reduced expert rejection by 7%. The complete system added 71ms latency per image while boosting accuracy by 0.057 compared to the baseline. Following the curation process, the final dataset comprises high-quality, clinically relevant images (see examples in Fig. 6).

Fig. 6.

Selected synthetic X-ray images.

Clinical Validation

Because exhaustive expert review of all 35,000 synthetic images was not practical, we performed a stratified blinded validation on a representative subset sampled from the post-generation quality-controlled dataset, with 10 images drawn from each pathology category (350 images in total). Three board-certified musculoskeletal radiologists independently reviewed these images under a blinded protocol. The evaluation protocol assessed three key aspects: anatomical plausibility using a 5-point Likert scale (where 1 indicates anatomically impossible and 5 indicates perfect anatomy), pathological authenticity through binary classification (authentic or non-authentic representation), and diagnostic confidence rated on a 0–10 scale (from no confidence to absolute confidence).

For pathological authenticity, the final label for each image was determined by majority voting among the three reviewers rather than by requiring total agreement. In other words, an image was considered authentic if at least two of the three reviewers classified it as authentic. For anatomical plausibility and diagnostic confidence, we report the mean score across the three reviewers.

The validation results, summarized in Table 4, demonstrated high anatomical fidelity with a mean plausibility score of 4.2 (standard deviation of 0.7). Pathological authenticity was accepted in 92% of cases (95% confidence interval: 89–94%). Diagnostic confidence averaged 8.7 for fracture cases compared to 7.2 for degenerative conditions, suggesting that fracture-related synthetic images were perceived by expert reviewers as more diagnostically convincing than degenerative cases.

Table 4.

Summary of Evaluation Metrics for Synthetic Cases by Board-Certified Radiologists.

Metric	Evaluation Method	Result
Anatomical Plausibility	5-point Likert scale (1 = Impossible, 5 = Perfect)	Mean score of 4.2 (±0.7)
Pathological Authenticity	Binary classification (Authentic/Non-authentic)	92% accepted as authentic (95% CI: 89–94%)
Diagnostic Confidence	0–10 scale (0 = No confidence, 10 = Absolute)	Fracture cases: 8.7
		Degenerative conditions: 7.2

Limitations

This study has several limitations. First, Hip-35 is a 2D hip radiograph/X-ray dataset rather than a volumetric CT dataset, and therefore does not provide multi-slice or 3D anatomical information. Second, although the synthetic images were reviewed by board-certified radiologists, this evaluation was performed on a stratified subset of post-generation quality-controlled images and should not be interpreted as establishing equivalence to full clinical diagnosis in all orthopedic scenarios. In addition, because the source cohort was retrospectively collected and de-identified, complete scanner-level acquisition parameters were not consistently available for all examinations. Accordingly, the dataset should be regarded as a supplementary resource for 2D radiographic research, rather than as a substitute for real clinical cohorts or for multi-view/volumetric imaging in diagnostic applications.

Usage Notes

The Hip-35 dataset can be used by other researchers for various purposes, including: (1) Benchmarking synthetic hip radiograph generation and post-generation quality-control pipelines; (2) Developing algorithms for classifying rare orthopedic conditions; (3) Studying the impact of synthetic data augmentation on model robustness. Researchers should be aware that the dataset, while diverse, primarily consists of synthetic images derived from a specific patient cohort. The provided quality control toolkit can be used to apply similar filtering to other medical imaging datasets.

Author contributions

P.W. developed the code, contributed to the labelling process, ran the experiments, and wrote the initial draft. Z.Z. and Y.Z. initiated the study, provided the dataset, validated the labels, and reviewed and edited the manuscript. J.L. revised the manuscript.

Funding

This research was funded by the Research Project of Chuzhou University (Grant No. 2025qd36).

Data availability

The Hip-35 dataset is deposited in the Hugging Face Hub (https://doi.org/10.57967/hf/8158)³⁴. The dataset comprises 35,000 synthetic radiographic images, which are all provided in PNG format. It is organized into 35 folders, with each folder corresponding to a specific pathology category and containing 1,000 images. The directory structure is well-organized, facilitating easy and rapid reuse.

Code availability

The Diffusion Mamba (DiM) model, which served as the backbone for synthetic image generation, is publicly available at https://github.com/tyshiwo1/DiM-DiffusionMamba/. The open-source code for our three-stage quality control pipeline, including chromatic artifact detection, multi-metric deblurring, and perceptual hashing deduplication, is available at https://github.com/Pu-Wang-alt/hip-joints/tree/main. All experiments were implemented using the PyTorch framework (version 2.0), which is publicly available at https://pytorch.org/.

Competing interests

The authors declare no competing interests.