Kashin-Beck disease diagnosis based on deep learning from hand X-ray images

https://doi.org/10.1016/j.cmpb.2020.105919Get rights and content

Highlights

  • An automated Kashin Beck Disease (KBD) diagnosis algorithm was proposed based on deep learning.

  • The method focuses on multi-features fusion for detection. The features are extracted from hand X-ray images by DCNN, including global and local features.

  • Global features represent structures of hand bones, and local features represent subtle edge information from critical regions of metaphysis based on the domain knowledge of KBD.

  • The model can effectively diagnose KBD and provide substantial benefits to reduce largescale screening costs and missed diagnosis rate.

Abstract

Background and objective

Kashin-Beck Disease (KBD) is a serious endemic bone disease leading to short stature. The early radiological examinations are crucial for potential patients. However, many children in rural China cannot be diagnosed in time due to the shortage of professional orthopedists. In this paper, an algorithm is developed to automatically screening KBD based on hand X-ray images of subjects, which can help the government reducing human resources investment and assisting the poor precisely.

Methods

The KBD diagnosis method focuses on multi-feature fusion for classification. Two kinds of features presented in X-ray images are extracted by a deep convolutional neural network (DCNN). One is the global features that represent shapes and structures of the whole hand bone. The other is local features that represent edge and texture information from critical regions of the metaphysis. The global features tend to sketch the major informative parts, whereas other fine local features can provide supplementary information. Then both kinds of features are combined and fed into the KBD classifier of a fully connected neural network (FCNN) to obtain diagnostic results.

Result

Our research team collected 960 samples in KBD endemic areas of Tibet from 2017 to 2018. The dataset contains 219 KBD positive images and 741 negative images. Experiments indicate that the method based on multi-feature achieves the best average accuracy and sensitivity rate of of 98.5% and 97.6% for diagnosis, which is 4.0% and 7.6% higher than the method with only the global features respectively.

Conclusions

The KBD diagnosis method shows that our proposed multi-feature fusion helps to achieve higher diagnosis performance and stability compared with only using global features for detection. The automated KBD diagnosis algorithm provides substantial benefits to reduce large-scale screening costs and missed diagnosis rate.

Introduction

Kashin-Beck Disease (KBD) is a serious endemic type of bone disease which leads to short stature. It primarily occurs in children ages 5-15 during bone-joint development, resulting in varying degrees of joint deformation and limited joint mobility [1]. It is mainly distributed from northeastern to southwestern China, Southeast Siberia, and North Korea.

There still is a high KBD incidence rate in rural areas. In recent years, KBD has been effectively controlled in China by implementing comprehensive preventive measures. However, owing to the shortage of professional orthopedists and medical equipment in rural areas, many potential patients do not get early radiological examinations and diagnoses leading to missing timely treatment. In 2013, there are 0.64 million patients with the KBD and 1.16 million at risk in 377 counties of 13 provinces or autonomous regions [2], and it remains a substantial threat to 22 million people in 2017. (KBD is a serious endemic type of bone disease and people who live in KBD endemic areas are under the threat.)

Early intervention can largely avoid KBD. Currently, hand X-ray examination for children is the best means for identifying KBD [1]. Although radiological examinations are accessible, KBD diagnosis is still a heavy task for orthopedists due to the requirement of long-term cultivation, shortage of professionals and the increasing number of monitoring areas. In addition, KBD positive X-ray signs [4] are difficult to identify early, which leads to missed diagnoses. Recently, with the popularity of medical Artificial Intelligence (AI) [5], [6], KBD diagnosis based on AI from hand X-ray images is a feasible way to alleviate the aforementioned conflicts. With deep learning namely AI, we can extract the wide variety of features [7] that are presented in medical images, including patterns, colors, values and shapes, to help doctors make accurate diagnose at a fast speed.

According to the Chinese national diagnostic criteria for KBD, “Diagnosis of KBD, WS/T 207-2001” [4], specific changes of metaphyseal zones are KBD positive X-ray signs. It can be observed as early as 5 years of age [1]. In practice, KBD monitoring is targeted at children aged 7–12 years old [3]. Typical KBD positive and normal hand X-ray images are shown in Fig. 1. Fig. 1a is the left hand from a 10-year-old boy with KBD positive signs: there are some metaphysis of hardening and interrupted signs, and small depressions waviness or serration changes (white arrows) in the zones of provisional metaphyseal calcification in the fingers (including index, middle and ring finger) [3]. Besides, zones of carpal bones are found crowded. As a contrast, Fig. 1b belongs to a 9-year-old healthy boy, of which the metaphysis is smooth. With deep learning, similar features can be extracted by DCNN [8] from hand X-ray images, which are used for identifying KBD. DCNN has demonstrated its ability to learn the most informative and typical features for image recognition [9], [10], [11].

The hand X-ray signs of Kashin-Beck Diseas are an integration of multiple features of two levels. One is the global features and the other is the local features. Both features are extracted by DCNN, while they come from the full hand image and key subgraphs (regions of the metaphysis, metacarpophalangeal joints, and carpals) respectively. The global features describe the comprehensive and distinguishable information of the hand image. In addition, it reflects characteristics of hand bones including shape, structure, thickness, density and KBD signs as mentioned above. The global features tend to grasp the most significant part but may miss other fine structures of hand bones which could provide supplementary information. For fine-grained detection of KBD, local features are obtained from critical regions. Two types of features are fused as the KBD features with concatenating.

In this paper, our proposed Kashin-Beck Disease diagnosis model aims to effectively locate metaphyseal regions of hand images and extract the aforementioned KBD features for diagnosis. The flow chart of the KBD diagnosis is shown in Fig. 2, which consists of three phases: preprocessing, feature extraction and KBD diagnosis. The most important phase is feature extraction including critical regions location, global and local features extraction, and feature fusion. To generate the global features and the local features, the preprocessed hand X-ray image, as well as the critical subgraphs, are fed into DCNN respectively. Then these two features are merged. Finally, the fusing KBD features are used for feature classification and KBD diagnosis.

In summary, (1) an algorithm is developed to automatically screening KBD. In related fields, KBD researchers have little attention to the automatic KBD diagnosis method. The traditional KBD diagnosis method is based on the manual features of hand X-ray images and it is extremely dependent on experienced professional doctors. Deep learning is introduced into KBD diagnosis and it is expected to eliminate the dependence of experienced orthopedists. (2) Our research team collected 960 hand X-ray images of subjects in KBD endemic areas of Tibet from 2017 to 2018, and built the KBD traceable digital database. (3) Experimented in the dataset, the results indicate that the method based on multi-feature achieves the best average accuracy and sensitivity rate of 98.5% and 97.6% for diagnosis. Meanwhile, the results are substantially consistent with the results of the experienced orthopedists from statistic test.

Section snippets

Overview of the KBD model

For an input image, it goes through three phases for diagnosis from left to right as shown in Fig. 3: image preprocessing, feature extraction, and Kashin-Beck Disease diagnosis. Firstly, the image is preprocessed. In the second phase, there are two branches of feature extraction. The preprocessed image is fed into ResNet [12] to extract the global features in the main branch. ResNet is a special DCNN and it is used as the backbone network for feature extraction in our model. In another branch,

Hand X-rag images dataset

The hand X-rag images dataset that we collected in Tibet from 2017 to 2018 contains about 960 images including 219 Kashin-Beck positive images and 741 negative images. These images are from the KBD subjects aged 7–12 years old. Each image is diagnosed and labeled by two KBD orthopedists and checked by four experienced assistants.

The dataset of hand X-ray images is divided into the training dataset and the testing dataset by sampling. The number of the train and the test is roughly 4:1, and the

Our contributions

Based on deep learning, we develop an automatic algorithm of large-scale Kashin-Beck Disease screening, which mainly focuses on global and local features of hand X-ray images. Our study involves three main parts: (1) the pre-train DCNN based on transfer learning [20] is applied to extract the features of hand images; (2) both global and local features are fused to diagnose KBD; (3) the traceable Hand X-rag images database is built and applied to evaluate the performance of the model.

The feature

Conclusion

In this paper, we propose a novel method for Kashin-Beck Disease diagnosis based on deep learning with hand X-ray images. We utilize the KBD features that combined with the global features extracted from the full hand image and auxiliary local features from critical regions to generate the class probability for diagnosing KBD. Experimental results show that our method has good performances on both basic KBD diagnosis and large-scale screening of KBD. Compared with the traditional manual method,

Declaration of Competing Interest

There are no known conflicts of interest associated with this publication. The financial support of this work will not influence its outcome.

Acknowledgments

This work was supported by the Fundamental Research Funds for the Central Universities (2020XD-A02-1). We would like to thank Zhichang Li, Shu Zhang, and Xiaobo Cheng for taking hand X-ray images. We would like to thank Hongqiang Gong, Shengcheng Zhao, and Min Guo for data collection. We would like to thank Qunwei Li, Ziyi Yang, Yudian Qiu, and Qiang Liu for KBD diagnosis and data annotations.

References (24)

  • X. Guo et al.

    Recent advances in the research of an endemic osteochondropathy in china: Kashin-beck disease

    Osteoarthr. Cartil.

    (2014)
  • X. Guo

    Diagnostic, clinical and radiological characteristics of Kashin-Beck disease in Shaanxi Province, PR china

    Int. Orthop.

    (2001)
  • Q. Song et al.

    The characteristics of positive and confusing hand X-ray signs in diagnosing Kashin-Beck disease in children in china

    Sci. Rep.

    (2018)
  • Chinese Ministry of Health M., Diagnosis of Kashin-Beck disease, (2010). WS/T 207-2010....
  • M. Bakator et al.

    Deep learning and medical diagnosis: a review of literature

    Multimodal Technol. Interact

    (2018)
  • J. De Fauw et al.

    Clinically applicable deep learning for diagnosis and referral in retinal disease

    Nat. Med.

    (2018)
  • R. Poplin et al.

    Predicting cardiovascular risk factors from retinal fundus photographs using deep learning

    Nat. Biomed. Eng.

    (2018)
  • I. Goodfellow et al.

    Deep Learning

    (2016)
  • A. Krizhevsky et al.

    Imagenet classification with deep convolutional neural networks

    Commun. ACM

    (2017)
  • K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale image recognition, arXiv preprint arXiv:...
  • V. Gulshan et al.

    Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs

    JAMA

    (2016)
  • K. He et al.

    Deep residual learning for image recognition

    CVPR

    (2016)
  • 1

    Jinyuan Dang, Hu Li, and Kai Niu equally contributed to this study and share the first authorship.

    View full text