Skip to main content
Journal of Imaging Informatics in Medicine logoLink to Journal of Imaging Informatics in Medicine
. 2025 Dec 15;39(5):3962–3971. doi: 10.1007/s10278-025-01783-1

Fine-Grained Differentiation-Based GNN for Auxiliary Diagnosis of Kashin-Beck Disease

Hu Li 2,#, Chendan Mu 1,#, Mengqi Li 1, Hui Li 2, Zixuan Wang 1, Kai Niu 1,3, Zhiqiang He 1,3,✉, Jianhao Lin 2
PMCID: PMC13694513  PMID: 41398140

Abstract

Kashin-Beck Disease (KBD), a localized osteoarthropathy, can lead to severe dwarfism and deformities, making early intervention crucial. Manual radiographic assessment of KBD can be challenging due to subtle changes of lesion features and overlapping features with other osteoarthropathies. Incorporating the radiological features of the foot is vital for diagnosing atypical KBD cases with unclear hand lesions. This paper employs graph neural networks (GNNs) to enhance feature differentiation in susceptible areas, distinguishing KBD lesions from similar conditions and aiding diagnosis. We identified seven regions in hand and ankle X-ray images based on prior knowledge as primary diagnostic focuses. Superpixel segmentation is applied within these regions, considering feature similarity and spatial proximity. A graph network is established with pixel blocks as nodes, highlighting lesion differences through edge weight variations. Finally, a convolutional neural network (CNN) extracts global features, enhancing focus on susceptible regions for precise KBD diagnosis. Experimental results show high diagnostic accuracy (90.97%) and recall (97.93%), with an overall F1-score of 93.97%, supporting the feasibility of large-scale automatic KBD diagnosis.

Keywords: Kashin-Beck Disease (KBD), Graph neural network (GNN), Superpixel segmentation

Introduction

Kashin-Beck Disease (KBD) is an endemic, deforming, chronic osteoarthropathy prevalent in regions such as southeastern Siberia, northern North Korea, and parts of China [2]. It primarily affects children and adolescents, causing symmetric degeneration and necrosis of epiphyseal cartilage, epiphyseal plate cartilage, and articular cartilage in the fingers, wrists, and ankles [4]. Early stages of KBD often lack obvious symptoms, but if untreated, it can result in motor disorders and severe deformities [3].

Radiographic examination of the hands and ankles is crucial for diagnosis. If the diagnosis is challenging to confirm through hand X-ray examination, an additional lateral view of the ankle joint should be obtained [1]. However, the similarity of KBD’s radiographic features to other osteoarthropathies, such as osteoarthritis, complicates accurate diagnosis [4].

The differential diagnosis between KBD and other bone and joint diseases is a critical aspect of KBD diagnosis [15]. Currently, the diagnosis of KBD primarily relies on manual radiographic interpretation by specialized doctors. The similarity it and other osteoarthropathies, as well as the diversity of radiographic manifestations among different degrees of it, increases the difficulty of manual interpretation and poses high demands on the doctors’ professionalism [3]. Additionally, manual interpretation is labor-intensive and time-consuming, which limits the feasibility of large-scale screening. Furthermore, the limited availability of specialized doctors, particularly in remote areas, further restricts access to timely diagnosis and treatment [13].

The advent of artificial intelligence has revolutionized medical diagnostics, yet research on KBD has predominantly focused on its etiology and prevention strategies [7–9]. There is limited research on diagnostic methods using deep learning [5]. In recent years, deep learning has played an important role in medical image processing and disease diagnosis assistance [16–18]. Existing studies have not fully addressed the similarity in radiographic features between KBD and other conditions, nor have they considered the ankle joint’s features [12, 13]. Therefore, utilizing deep learning to extract precise X-ray imaging features from both hands and ankles is critical for accurate diagnosis and large-scale screening of KBD [10, 11].

In this paper, we address the current research status and existing problems in auxiliary diagnostic methods by employing graph neural networks to model the relationships between superpixel blocks. We utilize superpixel segmentation and weighted capture of the relationships between superpixel blocks to distinguish lesions of KBD from other similar lesions, thereby aiding in the diagnosis of KBD. Specifically, the diagnostic criteria for KBD reveal that it primarily presents as multiple symmetrical depressions, sclerosis, and deformations on the bony articular surfaces of the metacarpophalangeal bones, wrist (ankle) joints, and metaphyseal calcification zones. Consequently, as depicted in Fig. 1, we identify the areas corresponding to the metacarpophalangeal bones, wrist, and ankle joints on the X-ray image as critical areas for diagnosing KBD. We employ clustering techniques for superpixel segmentation within these areas, constructing a graph network with pixel blocks as nodes. By weighted modeling of the graph network edges, we represent the relationships between pixel blocks, thereby facilitating the diagnosis of KBD while ruling out other related conditions.

Fig. 1.

Fig. 1

The regions prone to KBD lesions

Finally, by intensifying the focus on vulnerable regions within the global features extracted by CNNs, we achieve precise auxiliary diagnosis of KBD.

Experimental results demonstrate that our method achieves high diagnostic accuracy and recall, showing strong potential for reliable automatic diagnosis of KBD. It has high clinical application value and is of great significance for large-scale batch auxiliary diagnosis of KBD.

In summary, our specific contributions are as follows:

  1. We propose the fine-grained differentiation network (Fine-DiffNet), a novel approach that leverages superpixel segmentation and weighted graph modeling to capture relationships between pixel blocks. By constructing graph networks, Fine-DiffNet enhances attention to key regions within an image. This, coupled with the integration of multi-scale feature extraction, significantly improves the diagnosis of KBD. When evaluated on the KBD dataset collected by Peking University People’s Hospital, our method achieves superior performance in terms of both accuracy and recall compared to existing deep learning approaches.

  2. We identify the metacarpophalangeal joints, middle phalanges, wrist bones, and ankle joints as critical regions for diagnosing KBD. To analyze these areas, we apply K-means clustering to segment each region into distinct superpixel blocks, which are treated as nodes within a graph network. The edge weights between nodes represent the strength of connections between these superpixel blocks. By modeling the relationships and interactions among these blocks, our method effectively differentiates KBD from other bone and joint disorders.

  3. We adopt a combined diagnostic strategy focusing on both the hands and ankles to address the challenges of diagnosing atypical KBD. In the absence of typical hand manifestations, we place emphasis on ankle signs, while comprehensively analyzing the characteristics of multiple key regions in both the hands and feet to aid in the differential diagnosis of KBD. This approach broadens the diagnostic scope, enhancing the detection of atypical cases and improving overall diagnostic accuracy and reliability.

Related Works

In recent years, with the development of artificial intelligence, machine learning, and deep learning-based assisted diagnosis have played an increasingly important role in the medical field [5]. However, current research on KBD has primarily focused on exploring its etiology [6–9] and developing prevention and treatment strategies [10, 11]. There has been limited research on diagnostic methods for KBD. Lü et al. [14] explore and analyze the familial clustering and heritability of KBD in China through experiments. Only two studies [12, 13] have combined deep learning techniques to assist in the diagnosis of KBD, but these works have only roughly extracted radiographic features of the disease, leading to the omission of some detailed information and the need for improvement in diagnostic accuracy. Furthermore, they have not considered the radiographic features of the ankle joint, making it difficult to accurately diagnose cases with low distinguishability. Therefore, there is room for improvement in both accuracy and recall rates. Therefore, the utilization of deep learning technology to fully and precisely extract X-ray imaging features from the hands and ankles holds significant importance for the accurate and assisted diagnosis, as well as for large-scale and batch screening of KBD.

Method

This section provides a detailed introduction to the structure and design principles of Fine-DiffNet, which is used for the diagnosis of KBD. Firstly, we present an overall introduction to the framework of Fine-DiffNet, which mainly consists of three parts: a global feature extraction module, a module for manifesting differences in lesion characteristics, and an attention-based joint classification module. Subsequently, we will introduce the principles and functions of each part in separate subsections.

Framework of Fine-DiffNet

In this article, as shown in Fig. 2, we introduce a fine-grained differentiation network (Fine-DiffNet) that, building upon the global features extracted by CNNs, differentiates KBD lesion characteristics from osteoarthritis (OA) and other joint diseases by constructing GNNs within susceptible regions, thereby achieving more accurate automatic diagnosis of KBD. To clarify the radiographic basis of our diagnostic framework, Table 1 provides a detailed mapping between the WS/T 10026—2024 KBD diagnostic criteria and the corresponding processing steps in our method, as well as the extracted KBD-related features or regions. This table highlights the hallmark X-ray manifestations—such as metaphyseal and epiphyseal deformities, irregular metaphyseal lines, and phalangeal end changes—that are captured and analyzed by the proposed Fine-DiffNet.

Fig. 2.

Fig. 2

Overall architecture of fine-grained differentiation network (Fine-DiffNet)

Table 1.

Mapping between KBD diagnostic criteria, processing step in the method, and extracted KBD-related features or regions

Corresponding clause and description in diagnostic criteria Processing step in the method Extracted KBD-related features or regions
Corresponding clauses: B.2.1.1–B.2.1.4 (hand); B.2.2.1–B.2.2.3 (foot) Prior-knowledge-based identification of seven diagnostic regions Hand: metaphysis, epiphysis, epiphyseal plate of phalanges, carpal bones; Foot: talus and calcaneus
Assists identification of: B.2.1.1.1: trabecular disorganization, hardened depressions; B.2.2.2: talar articular surface depression, calcaneus shortening Fine-grained extraction via superpixel segmentation Helps detect features such as sclerosis, depression, and bone defects used in grading (“+,” “++,” “+++”)
Salient connections reveal: B.2.1.2.2: marginal bone defect, cystic changes; B.2.1.3.2/3: epiphyseal deformation, slanting, defects; B.2.2.3: talar collapse, marginal loss Graph neural network construction Detects significant inter-region shape/density changes such as deformity, cystic change, collapse
Extracts compound features across modalities: 5.1 When hand radiographs are inconclusive, additional ankle lateral views are acquired for comprehensive assessment CNN-based global feature extraction with enhanced focus on key regions Combines hand and ankle radiographic information to diagnose atypical KBD when hand X-ray alone is insufficient

The Fine-DiffNet mainly consists of three components: a global feature extraction module, a module for manifesting differences in lesion characteristics, and an attention-based joint classification module.

The global feature extraction module utilizes CNNs to extract multi-scale features from images and obtain global feature vectors. Based on prior knowledge, seven susceptible regions are identified, and the global features of each susceptible region are obtained by extracting multi-scale features from each region. Then, concatenate all global feature maps to obtain the final global feature map. The module for manifesting differences in lesion characteristics first uses clustering methods to segment each susceptible area into different pixel blocks, extracts features from each pixel block using CNNs, and then constructs GNNs in the susceptible area using pixel blocks as nodes to amplify lesion features and highlight differences between similar features. On the basis of global multi-scale features, the attention-based joint classification module enhances attention to the differential features of lesions in susceptible areas, and combines the weight allocation of different feature maps to obtain classification results, achieving automatic diagnosis of KBD.

The framework proposed in this paper, Fine-DiffNet, is an end-to-end trained model that takes concatenated images of the hands and ankles as input. During the training process, it combines the results obtained from the regression of features at different levels to derive a probability value for predicting KBD. Specifically, for each input sample, the model outputs a predicted probability y^i that the sample belongs to the class of having KBD. The true label for each sample is denoted as yi, where yi=1 indicates the presence of KBD and yi=0 indicates its absence. To evaluate the performance of the model, we calculate the loss using the predicted result and the true label. The loss function adopted in this work is the binary cross-entropy loss, defined as follows:

Loss=-1N∑i=1Nyilog(y^i)+(1-yi)log(1-y^i) 1

where N is the total number of samples in the training set. This loss function quantifies the discrepancy between the predicted probability distribution and the true binary distribution, enabling the model to iteratively adjust its parameters during training in order to minimize the loss and improve its predictive accuracy for KBD.

Global Feature Extraction Module

The global feature extraction module primarily focuses on extracting global features from the image and selecting susceptible regions based on prior knowledge. It performs local feature extraction at a coarse-grained level within these susceptible regions, achieving a shallow level of feature refinement. Subsequently, the global features and the coarse-grained local features are concatenated to obtain a final global feature map. This process ensures that both overall and specific aspects of the image are captured, laying the foundation for more accurate disease diagnosis.

The global feature extraction module utilizes ResNet50 as the backbone network, where the core consists of convolutional layers performing feature extraction through convolution operations. Initially, the module extracts features from the combined input image of the hand and ankle using ResNet50, with an input size of 512×512×3, resulting in a global feature map of size 32×32×2048. According to prior medical knowledge, as shown in Fig. 1, KBD lesions commonly occur in the middle and proximal phalanges of the index, middle, and ring fingers, as well as the terminal phalanx of the thumb and the ankle. The module crops the image to obtain the combined regions of the middle phalanges of the index, middle, and ring fingers, the combined regions of the proximal phalanges of the index, middle, and ring fingers and the thumb, and the ankle region. Each cropped region is resized to a fixed input size of 512×512×3 to fit ResNet50’s input requirements. Following this, ResNet50 is employed to extract features from each of these susceptible regions, producing seven sets of coarse-grained local feature maps, corresponding to the different regions of interest. Subsequently, we use ResNet50 to extract features from each of these susceptible regions, obtaining seven sets of coarse-grained local feature maps corresponding to different susceptible regions. Let I denote the input image, and Ri represent the i-th susceptible region cropped from I. The feature extraction for region Ri using ResNet50 can be formulated as:

Fi=ReLUBNWi∗I[Ri]+bi 2

where Fi is the feature map corresponding to region Ri, Wi and bi are the learnable weight and bias parameters, BN denotes batch normalization, and ∗ represents the convolution operation. Subsequently, the final feature map is constructed by concatenating the global feature map with the local feature maps extracted from the different susceptible regions:

Ffinal=ConcatFglobal,F1,F2,⋯,Fn 3

where Fglobal represents the global feature map extracted from the entire input image, F1,F2,⋯,Fn denote the local feature maps corresponding to the different lesion-prone regions, and Concat refers to the concatenation operation along the feature dimension.

This feature map preserves the global features of the original image while comprehensively utilizing the coarse-grained local features of the susceptible regions. By combining multi-scale features, it achieves shallow-level fine-grained feature processing.

Module for Manifesting Differences in Lesion Characteristics

Given that KBD’s susceptible regions exhibit subtle, fine-scale changes—such as trabecular disorganization, hardened depressions, marginal bone defects, cystic alterations, and collapse—coarse-grained CNN features alone cannot capture these patterns. Therefore, it is necessary to employ a more fine-grained processing of the features in these susceptible regions to mine the detailed texture, density, and shape information. The features in these susceptible regions are extracted from both hand and ankle x-rays, according to the diagnosis criteria.

As shown in Fig. 3, in this module, we first apply the Simple Linear Iterative Clustering (SLIC) algorithm to the seven susceptible regions obtained from the coarse-grained feature extraction module. This algorithm divides each susceptible region into multiple superpixels by initializing and selecting cluster centers, assigning labels to each pixel, and measuring the color and spatial distances between pixels and cluster centers. Specifically, the process can be expressed as:

Superpixels=SLIC(I[Ri],K,m) 4

where I[Ri] is the image of the i-th susceptible region, K is the number of superpixels to generate, m is the compactness parameter, and SLIC(·) represents the SLIC algorithm.

Fig. 3.

Fig. 3

Graph network construction

We then perform padding and resizing (if necessary) on the superpixels and extract features from each of them using ResNet50, enabling a more detailed sampling and processing of local features. This step can be conceptually represented as:

Fspj=ResNet50(PadAndResize(Superpixelj)) 5

where Superpixelj is the j-th superpixel, PadAndResize(·) represents the combined padding and resizing operation (which may be implemented as separate steps), and Fspj is the extracted feature from that superpixel.

Finally, for each susceptible region, we construct a Graph Convolutional Neural Network (GCNN) with superpixels as nodes to capture the intricate relationships between them. This allows for a more precise extraction and characterization of the local features in the susceptible regions. The GCNN process can be conceptually described as:

GCNN\_Output=GCNN({Fsp1,Fsp2,…,FspK},A) 6

where {Fsp1,Fsp2,…,FspK} is the set of features extracted from all superpixels, A represents the adjacency matrix or some other form of encoding the relationships between nodes, and GCNN(·) is the function representing the graph convolutional neural network.

By leveraging this fine-grained feature extraction, we can distinguish the unique lesion features of KBD from the similar features of other musculoskeletal diseases, thereby enhancing the accuracy of KBD diagnosis through fine-grained feature processing.

Attention-Based Joint Classification Module

The proposed method integrates diagnostic information from both global and local perspectives to produce the final prediction. The global feature extraction module captures overall morphological features from hand–ankle X-rays, while the lesion feature differentiation module analyzes seven susceptible local regions identified from the global feature maps. Within the CNN branch, multiple feature maps are concatenated along the channel dimension to enrich the semantic representation.

For each of these regions, superpixel segmentation is first applied to form local structural units. A GCNN is then constructed for each region, treating the superpixels as nodes and their spatial or feature-based relationships as edges. Each GCNN performs an independent classification to determine whether the corresponding region exhibits lesion features related to KBD. This enables fine-grained modeling of localized pathological changes that may not be captured by global CNN features.

The outputs of the GNN and CNN modules are combined at the decision level according to a logical OR rule:

Prediction=1,if any of{R1,R2,...,R7,G}=1,0,otherwise.

Here, G represents the global classification result from the CNN, and Ri denotes the output of the GCNN corresponding to the i-th local region. Thus, if any of the local or global classifiers outputs a positive prediction, the overall sample is labeled as KBD.

This decision-level integration strategy enhances diagnostic sensitivity by allowing lesion evidence from any region to independently contribute to the final prediction. Figure 2 also illustrates this integration logic.

Experiment

Dataset

We collected and processed hand and ankle X-ray images from the KBD epidemic area in Heilongjiang Province, China to construct a joint diagnostic dataset for KBD, to validate and evaluate the performance of our algorithm. We screened people from some KBD-endemic villages in Heilongjiang Province and collected X-ray images of residents’ hands and ankles. The dataset contains 207 residents (108 males, 99 females, mean age 59.9 ± 7.6 years) from KBD-endemic areas of Heilongjiang Province, including 126 KBD patients (61 males, 65 females, mean age 59.0 ± 6.4 years) and 81 healthy controls. According to the Chinese National Diagnostic Criteria for KBD, patients were stratified into three grades: 69 cases (54.8%) were classified as Grade I, 39 cases (30.9%) as Grade II, and 18 cases (14.3%) as Grade III. The disease markers for each X-ray (a combination of ankle and hand X-rays for each person) are marked by a professional doctor.

All procedures performed in this study complied with relevant laws and institutional guidelines. The study protocol was approved by the Ethics Committee of Peking University People’s Hospital and authorized by the Health Department of the Tibet Autonomous Region (Document No. ZangWeiSanBingHan [2018] No. 248). All radiographs were fully anonymized to ensure patient privacy, and informed consent was obtained from all participants prior to inclusion.

In order to facilitate the training and evaluation of neural network models, we will preprocess the collected images. Firstly, we will complete the missing hand image data by symmetrically flipping the known hand images. Then, we rotate the hand and ankle images to unify the direction and angle of the images in the dataset. Finally, we concatenated the data from our hands and ankles into one image, and annotated each image according to the doctor’s annotations, constructing a final dataset containing 207 images.

Considering the small sample size and imbalanced categories in the dataset, we adopted a stratified 5-fold cross-validation strategy to ensure that each fold maintained a similar class ratio between each level of KBD and Non-KBD cases. The final performance was reported as the average result of stratified 5-fold cross-validation repeated 5 times (25 test runs), which provides a more reliable estimate of model generalization and stability.

Implementation Details

The proposed method was implemented using PyTorch 2.0.1 and Python 3.9. The image size of the input network is 512 * 512, the optimizer uses Adam, the initial learning rate is 0.001, the loss function is cross-entropy loss, 200 epochs were trained, and the batch size is 32.

We performed data augmentations on datasets, including random center cropping at a scale of 0.8 to 1.0, random Gaussian blur with kernel sizes ranging from 5 to 30, random Contrast Limited Adaptive Histogram Equalization (CLAHE) processing, and random horizontal flipping.

The GNN network comprises four graph convolution layers with feature dimensions progressively reducing from the initial hidden dimension to 512, 256, 128, and finally 64 features per node. The adjacency matrix was constructed based on both spatial proximity and intensity similarity between superpixel regions, where edge weights were computed using the region adjacency graph (RAG) algorithm. For graph-level classification, we employed global mean pooling to aggregate node features, followed by two fully connected layers (64→32→2) with ReLU activation and dropout regularization (p=0.5).

Evaluating Indicator

We use commonly used evaluation metrics for image classification, such as Accuracy, Precision, Recall, and F1 score, as the evaluation metrics for this article. Due to the relative importance of recall and accuracy in medical tasks, we use recall and accuracy as the main indicators.

Accuracy is the most intuitive evaluation metric, which reflects the overall performance of a classifier on all samples.

Recall, also known as sensitivity, is defined as the proportion of true positives that are correctly identified as positive out of all true positives.

Precision focuses on the proportion of positive samples that are truly positive out of all samples classified as positive.

F1-score is the harmonic mean of Precision and Recall, which provides a single score that balances both Precision and Recall.

Results

To evaluate the performance of our method, we conducted experiments on the aforementioned dataset, employing accuracy, precision, recall, and F1-score as the four key metrics to assess the model’s performance. Given that the current diagnosis of KBD still relies primarily on manual inspection, we compared our method with commonly used image classification models to demonstrate its competitiveness. Specifically, ResNet50 and Vision Transformer (ViT) were adopted as baseline models. ResNet50 is a convolutional neural network that excels at capturing local image features through its hierarchical structure, while ViT utilizes a self-attention mechanism to model global dependencies across the entire image. In contrast to these image-level classifiers, Fine-DiffNet is designed to perform fine-grained, region-aware analysis through the integration of superpixel segmentation and graph-based structural modeling, thereby enabling more precise localization of subtle KBD-related lesions. The experimental results are presented in Table 2.

Table 2.

Performance comparison of different models with statistical significance analysis

Model Accuracy (%) Precision (%) Recall (%) F1-Score (%) p-value
ResNet50 81.30 ± 6.2 89.23 ± 6.0 76.80 ± 6.8 87.99 ± 6.1 <0.01
ViT 84.15 ± 6.0 91.14 ± 6.4 81.18 ± 6.7 91.12 ± 6.3 <0.05
Deep learning 89.02 ± 5.8 93.79 ± 6.1 91.16 ± 6.0 92.30 ± 5.7 <0.05
Our method 90.97 ± 5.5 90.44 ± 5.8 97.93 ± 5.1 93.97 ± 5.4 –

Values are shown as mean ± standard deviation. Statistical significance is calculated using McNemar’s test with respect to our method

From the results presented above, it is evident that our method exhibits significant improvements in accuracy and recall compared to existing baseline models. Specifically, when compared to the convolutional neural network classification model, ResNet50, our method achieves over 9% higher accuracy and more than 20% higher recall. Furthermore, in comparison to the attention-based ViT model, our method outperforms it by approximately 6% in accuracy and around 16% in recall.

The substantial increase in accuracy demonstrates that our method possesses strong overall classification capabilities. Meanwhile, the notable enhancement in recall indicates that our method performs exceptionally well in identifying disease samples, possessing a robust ability to detect patients. This is of utmost importance for the screening of KBD, as early detection can lead to prompt intervention and improved patient outcomes.

However, as can be seen from Table 2, compared to the baseline model, our method exhibits weaker performance in terms of precision. This is because the baseline model is more proficient in classifying normal samples while being less capable of classifying diseased samples. While we aim to achieve improvements in every evaluation metric, there exists a balance between precision and recall, where enhancing one metric may lead to a decrease in the other. Therefore, we introduce the F1-score, which combines precision and recall, to provide a more objective evaluation of our model. From the results, it is evident that in terms of F1-score, our model outperforms ResNet50 by nearly 6% and outperforms ViT by approximately 3%, demonstrating a significant classification advantage.

To further validate the competitiveness of our method, we applied existing deep learning approaches for automatic KBD diagnosis to our dataset and evaluated them using the same evaluation metrics. The experimental results are shown in Table 2. As can be observed from the results, our method exhibits significant improvements in terms of accuracy, F1-score, and especially recall, indicating that our approach outperforms the original deep learning algorithms in the task of automatic KBD diagnosis. Statistical analysis based on McNemar’s test across five folds showed that the proposed method achieved significantly higher performance than the baseline models (p<0.05), confirming the robustness of the improvement despite the limited dataset size.

As shown in Table 3, our method demonstrates superior performance across different KBD severity levels compared with the baseline models. The most notable improvement is observed in Level I cases, where the accuracy increases from 85.55% in the previous deep learning approach to 94.15%, representing an improvement of approximately 8.6 percentage points. This indicates that the proposed fine-grained differentiation technique effectively enhances the identification of early-stage KBD, which is typically the most challenging stage to diagnose due to subtle joint deformities and less distinctive imaging features.

Table 3.

Classification accuracy of different models on the test set for Non-KBD and KBD severity levels (I-III) in 5-fold cross validation

Model Non-KBD (%) Level I (%) Level II (%) Level III (%)
ResNet50 75.29 75.36 94.87 100.00
ViT 81.15 78.50 95.62 100.00
Deep learning 87.20 85.55 99.20 100.00
Our method 84.70 94.15 100.00 100.00

For moderate and severe cases (Levels II and III), all models achieve nearly perfect classification performance, suggesting that these stages are easier to distinguish because of their more apparent structural changes. In contrast, the lower accuracy in Non-KBD classification (84.70%) reflects the increased sensitivity of our model toward detecting potential disease features, even in borderline or atypical cases. Overall, the results confirm that the proposed approach substantially improves early-stage diagnostic capability while maintaining excellent performance for advanced KBD cases, providing a more balanced and clinically meaningful diagnostic framework.

Despite these promising results, this study has several limitations. The dataset used is relatively small, which may limit the generalizability of the proposed model and contribute to the observed variability in evaluation metrics. This effect can also be reflected in the relatively large standard deviations in Table 2, indicating that the small sample size has led to greater fluctuations in model performance across folds. A limited number of samples also increases the risk of overfitting, as the model may capture dataset-specific features rather than generalizable diagnostic patterns.

Ablation Study

To conclusively validate the contribution of each key component in our proposed Fine-DiffNet, we conducted a comprehensive ablation study. The study was designed to isolate and evaluate the effects of: (1) the graph neural network (GNN) module for fine-grained feature differentiation, and (2) the incorporation of ankle joint radiographs. All experiments were conducted under the same settings as described in the “Implementation Details” section.

As shown in Table 4, the results of the ablation study demonstrated the incremental contribution of each component within Fine-DiffNet. The baseline configuration, which utilized only hand radiographs without incorporating the GNN module, achieved an accuracy of 85.11% and a recall of 84.21%. When ankle joint radiographs were incorporated, recall was improved to 90.20%, further validating the importance of multi-joint analysis in detecting atypical KBD cases. When only the GNN module was applied, performance was significantly enhanced, achieving 88.90% accuracy and 92.86% recall, thereby confirming its capability to model fine-grained structural relationships among diagnostic regions. The complete framework, which integrated both components, achieved the optimal performance, thereby demonstrating the synergistic effects between multi-joint data integration and graph-based feature differentiation in KBD diagnosis.

Table 4.

Ablation study of Fine-DiffNet components

Component Choice
GNN module × × ✓ ✓
Ankle joint × ✓ × ✓
Accuracy 85.11 87.46 88.90 90.97
Recall 84.21 90.20 92.86 97.93

Disscussion

Our proposed Fine-DiffNet framework builds upon and extends recent advances in radiograph-based quantitative pipelines for skeletal disorders. Deep learning models have demonstrated reliable automated readouts from routine hand radiographs, as evidenced by studies on pediatric bone mineral density estimation [19] and osteoporosis screening [20], underscoring the capability of deep learning to extract quantitative biomarkers from standard X-ray images.

In parallel, task-adjacent research on hand radiograph analysis, including bone age assessment approaches based on Vision Transformers and hybrid CNN–ViT architectures [21, 22], has shown that jointly modeling global and local skeletal relationships can substantially improve predictive accuracy. These works collectively motivate our adoption of multi-scale feature extraction and attention mechanisms.

In contrast to bone age assessment, which evaluates global skeletal maturation, KBD diagnosis requires fine-grained, region-specific characterization of subtle pathological alterations. In the field of KBD-specific deep learning, recent studies [23, 24] have investigated transfer learning on metaphyseal signs and multi-joint fusion strategies, achieving high accuracy on large-scale datasets. However, such methods primarily perform image-level classification and often overlook localized, sub-structural lesion patterns. In contrast, Fine-DiffNet explicitly incorporates clinical prior knowledge to delineate seven anatomically susceptible regions and applies superpixel segmentation combined with a graph neural network to model intra-regional structural dependencies. This design facilitates more granular and interpretable feature representations tailored to KBD’s radiographic hallmarks. Moreover, while prior works relied on hand or hand–knee image pairs, our framework adopts the clinically recommended hand–ankle joint assessment approach, ensuring consistency with established diagnostic standards and enhancing robustness in atypical cases where hand manifestations are ambiguous. By integrating graph-based fine-grained modeling with multi-region attention, Fine-DiffNet not only complements existing CNN–ViT systems but also aligns with established radiographic assessment practices for KBD.

Conclusion

In this paper, we propose a fine-grained differentiation network (Fine-DiffNet based on CNNs and GNNs to effectively extract features from X-ray images of hands and ankles for accurate automatic screening of KBD. We extract features from three dimensions: global, local, and fine-grained. Global features are directly extracted using CNNs. Local features are obtained through the detection and feature extraction of vulnerable regions. Fine-grained features are captured by segmenting the vulnerable regions into superpixels and modeling the detailed differences between samples using a graph network. Finally, the multi-level features are combined to achieve automatic diagnosis of KBD through a comprehensive classification. Experimental results on the dataset demonstrate that our method achieves high accuracy, significantly outperforms common image classification baseline models in terms of recall and F1-score, and still shows prominent performance compared to existing deep learning models for automatic KBD diagnosis. Our approach holds significant clinical application value and is of great importance for large-scale assisted diagnosis of KBD.

Author Contributions

Hu Li and Chendan Mu contributed equally to this work. Conceptualization: Zhiqiang He, Jianhao Lin. Methodology: Chendan Mu, Hu Li, Mengqi Li. Data collection: Hu Li, Hui Li, Zixuan Wang. Data analysis and interpretation: Chendan Mu, Mengqi Li, Zixuan Wang, Kai Niu. Writing—original draft preparation: Chendan Mu, Mengqi Li. Writing—review and editing: Zhiqiang He, Kai Niu. Supervision: Zhiqiang He, Jianhao Lin

Funding

This research was funded by the National Key Research and Development Program of China (2022YFC2503103) Natural Science Foundation of Beijing Municipality (L244019).

Data Availability

The data supporting this study are not publicly available due to ethical restrictions. The clinical data of participants were only used for the purpose of this research project and cannot be shared publicly in accordance with the informed consent agreements signed by the patients and the institutional review board regulations.

Declarations

Ethics Approval

All procedures performed in this study were conducted in accordance with relevant laws and institutional guidelines, and were approved by the Ethics Committee of Peking University People’s Hospital on [Approval Date in 2018] (Approval No. ZangWeiSanBingHan [2018] No. 248).

Consent to Participate

Radiographic images used in this study were fully anonymized to ensure the privacy of the patients. Informed consent was obtained from all individuals whose data were included in the study. The study complied with the ethical standards of the institutional and national research committees and with the 1964 Helsinki Declaration and its later amendments.

Consent for Publication

Not applicable. Radiographic images were fully anonymized to ensure patient privacy, and thus consent to publish is not required.

Competing Interests

The authors declare no competing interests.

Footnotes

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Hu Li and Chendan Mu contributed equally to this work.

References

  • 1.National Health Commission of the People’s Republic of China: Chinese WS/T Diagnostic Criteria for Kashin-Beck Disease (2007-2010).
  • 2.Wang K, Yu J, Liu H, Liu Y, Liu N, Cao Y, Zhang X, Sun D: Endemic Kashin–Beck disease: A food-sourced osteoarthropathy. Semin Arthritis Rheum 50(2):366–372, 2020. [DOI] [PubMed] [Google Scholar]
  • 3.Guo X, Ma WJ, Zhang F, Ren FL, Qu CJ, Lammi MJ: Recent advances in the research of an endemic osteochondropathy in China: Kashin–Beck disease. Osteoarthritis Cartilage 22(11):1774–1783, 2014. [DOI] [PubMed] [Google Scholar]
  • 4.Wang X, Ning Y, Liu A, Qi X, Liu M, Zhang P, Guo X: The systematic review and meta-analysis of X-ray detective rate of Kashin–Beck disease from 1992 to 2016. BMC Musculoskelet Disord 20:1–9, 2019. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Rana M, Bhushan M: Machine learning and deep learning approach for medical image analysis: diagnosis to detection. Multimed Tools Appl 82(17):26731–26769, 2023. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6.Wang J, Zhao S, Yang L, Gong H, Li H, Nima C: Assessing the health loss from Kashin–Beck disease and its relationship with environmental selenium in Qamdo district of Tibet, China. Int J Environ Res Public Health 18(1):11, 2021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Wang X, Ning Y, Li C, Gong Y, Huang R, Hu M, Poulet B, Xu K, Zhao G, Zhou R, et al: Alterations in the gut microbiota and metabolite profiles of patients with Kashin–Beck disease, an endemic osteoarthritis in China. Cell Death Dis 12(11):1015, 2021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Lei J, Amhare AF, Wang L, Lv Y, Deng H, Gao H, Guo X, Han J, Lammi MJ: Proteomic analysis of knee cartilage reveals potential signaling pathways in pathological mechanism of Kashin–Beck disease compared with osteoarthritis. Sci Rep 10(1):6824, 2020. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9.Fu G, Chen X, Qi M, Du X, Xia Z, Liu Q, Sun N, Shi C, Zhang R: Status and potential diagnostic roles of essential trace elements in Kashin–Beck disease patients. J Trace Elem Med Biol 69:126880, 2022. [DOI] [PubMed] [Google Scholar]
  • 10.Tang X, Zhou ZK, Liu G, Shen B, Kang PD, Li J, Li Q, Pei FX: The long-term efficacy and tolerability of hyaluronic acid in adult patients with Kashin–Beck disease of the knee. Clin Rheumatol 34:151–156, 2015. [DOI] [PubMed] [Google Scholar]
  • 11.Yue J, Yang M, Yi S, Dong B, Li W, Yang Z, Lu J, Zhang R, Yong J: Chondroitin sulfate and/or glucosamine hydrochloride for Kashin–Beck disease: a cluster-randomized, placebo-controlled study. Osteoarthritis Cartilage 20(7):622–629, 2012. [DOI] [PubMed] [Google Scholar]
  • 12.Qiao Y, Li H, Niu K, Wang L, Lin J, He Z: A method for Kashin–Beck disease auxiliary diagnosis based on the features in regions of the potential lesion. Med Phys 50(10):6259–6268, 2023. [DOI] [PubMed] [Google Scholar]
  • 13.Dang J, Li H, Niu K, Xu Z, Lin J, He Z: Kashin–Beck disease diagnosis based on deep learning from hand X-ray images. Comput Methods Programs Biomed 200:105919, 2021. [DOI] [PubMed] [Google Scholar]
  • 14.Lü AL, Guo X, Aisha MMT, Shi XW, Zhang YZ, Zhang YY: Kashin–Beck disease and Sayiwak disease in China: prevalence and a comparison of the clinical manifestations, familial aggregation, and heritability. Bone 48(2):347–353, 2011. [DOI] [PubMed] [Google Scholar]
  • 15.Li FS, Duan YJ, Yan SJ, Guan JY, Zou LM, Wei FC, Mong LY, Li L, Li SY: Presenile (early ageing) changes in tissues of Kaschin–Beck disease and its pathogenetic significance. Mech Ageing Dev 54(2):103–120, 1990. [DOI] [PubMed] [Google Scholar]
  • 16.Tiwari A, Poduval M, Bagaria V: Evaluation of artificial intelligence models for osteoarthritis of the knee using deep learning algorithms for orthopedic radiographs. World J Orthop 13(6):603, 2022. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Yao B, Jin L, Hu J, Liu Y, Yan Y, Li Q, Lu Y: Noise-imitation learning: unpaired speckle noise reduction for optical coherence tomography. Phys Med Biol, 10.1088/1361-6560/ad3e79, 2024. [DOI] [PubMed]
  • 18.Dong Q, Ullah MN, Innes DR, Watkins RD, Chang C, Zou SJ, Groll A, Sacco I, Chinn G, Levin CS: PETcoil: first results from a second-generation RF-penetrable TOF-PET brain insert for simultaneous PET/MRI. Phys Med Biol, 10.1088/1361-6560/ad3d45, 2024. [DOI] [PMC free article] [PubMed]
  • 19.Zhao H, Zhang Y, Zhang W, et al: An automatic deep learning-based bone mineral density measurement method using X-ray images of children. Quant Imaging Med Surg 15(3):2481–2493, 2025. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Ho CS, Fan TY, Kuo CF, et al: HarDNet-based deep learning model for osteoporosis screening and bone mineral density inference from hand radiographs. Bone 190:117317, 2025. [DOI] [PubMed]
  • 21.Wu J, Mi Q, Zhang Y, et al: SVTNet: Automatic bone age assessment network based on TW3 method and vision transformer. Int J Imaging Syst Technol 34(2):e22990, 2024. [Google Scholar]
  • 22.Spampinato C, Palazzo S, Giordano D, et al: Deep learning for automated skeletal bone age assessment in X-ray images. Med Image Anal 36:41–51, 2017. [DOI] [PubMed] [Google Scholar]
  • 23.Li G, Zhang H, Chen P, et al: A comparison of transfer learning metaphyseal sign diagnostic models for Kashin–Beck disease based on X-rays of children’s hands. Cureus 17(1), 2025. [DOI] [PMC free article] [PubMed]
  • 24.Liu Z, Yang B, Li M, et al: Deep learning-based X-ray diagnostic method for Kashin–Beck disease. In: Proceedings of the 2024 8th International Conference on Digital Signal Processing. pp. 212–218, 2024.

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Data Availability Statement

The data supporting this study are not publicly available due to ethical restrictions. The clinical data of participants were only used for the purpose of this research project and cannot be shared publicly in accordance with the informed consent agreements signed by the patients and the institutional review board regulations.


Articles from Journal of Imaging Informatics in Medicine are provided here courtesy of Springer

RESOURCES