Computerized assisted evaluation system for canine cardiomegaly via key points detection with deep learning

https://doi.org/10.1016/j.prevetmed.2021.105399Get rights and content

Abstract

Cardiomegaly is the main imaging finding for canine heart diseases. There are many advances in the field of medical diagnosing based on imaging with deep learning for human being. However there are also increasing realization of the potential of using deep learning in veterinary medicine. We reported a clinically applicable assisted platform for diagnosing the canine cardiomegaly with deep learning. VHS (vertebral heart score) is a measuring method used for the heart size of a dog. The concrete value of VHS is calculated with the relative position of 16 key points detected by the system, and this result is then combined with VHS reference range of all dog breeds to assist in the evaluation of the canine cardiomegaly. We adopted HRNet (high resolution network) to detect 16 key points (12 and four key points located on vertebra and heart respectively) in 2274 lateral X-ray images (training and validation datasets) of dogs, the model was then used to detect the key points in external testing dataset (396 images), the AP (average performance) for key point detection reach 86.4 %. Then we applied an additional post processing procedure to correct the output of HRNets so that the AP reaches 90.9 %. This result signifies that this system can effectively assist the evaluation of canine cardiomegaly in a real clinical scenario.

Introduction

Large dogs over five years old would be diagnosed as heart disease when their symptoms are obvious enough to be easily realized by their owners, and cardiomegaly is the main imaging finding for diagnosing canine heart diseases. Radiological differences include the size, contour, position and density of great vessels between the normal and abnormal dogs. Canine cardiomegaly can be evaluated by empirical evaluation, comparative evaluation and measurement evaluation methods. X-ray is frequently used to differentiate this finding by experienced doctor and the evaluation of canine cardiomegaly largely depends on their experience and subjective computation. These methods have their own advantages and disadvantages in clinical practice.

For measurement evaluation method, veterinarians usually use VSS (Vertebral Scale System) (James, 2000) to precisely measure the distance between different anatomical parts clinically. This method is the key in evaluating this finding by obtaining the relative difference between the sizes of the hearts of normal and abnormal dogs. Though canine cardiomegaly is common in dogs and the diagnostic criteria is unified, the image reading is time-consuming and subjective. The intersection points between the first to the the twelfth ribs and the spine are marked with 12 key points (points 1–12). The remaining four key points are located at the bifurcation of main trachea (point 13), the apex of heart (point 14), the cranial edge of heart (point 15), and the intersection point between the caudal margin of heart and the caudal vena cava (point 16) respectively. All these key points are summarized in Fig. 1. The long axis is from point 13–14; while the short axis is from point 15–16 (Fig. 1).

VHS measures the relative size of heart on the basis of the size of intercostal space. Deep learning is a type of powerful data analyzing tool which has been successfully used in medical field for human being (Yang et al., 2019; Li et al., 2020), including diagnosis with medical imaging (Wang et al., 2017; Zhang et al., 2020b) and video (Zhang et al., 2020a). It’s power in veterinary medicine and animal related fields is also exploited by many researchers (Banzato et al., 2018; Bleuer-Elsner et al., 2019; Bradley et al., 2019; Kim et al., 2019; Gergely et al., 2020; Patel et al., 2020; Romero et al., 2020; Smith et al., 2020; Trachtman et al., 2020; Wang et al., 2020). Moreover, there is an empirical formula that could assist doctors in evaluating whether a dog is suffering from canine cardiomegaly. The premise is that veterinarians should precisely discern each anatomical parts and key positions on vertebra and heart. Therefore, we attempt to apply key points detection to help veterinarians diagnose canine cardiomegaly using deep learning.

Burti et al. applied convolutional neural network (CNN) to directly diagnose this disease, all images were classified as having a normal cardiac silhouette (No-VHS-Cardiomegaly) or an enlarged cardiac silhouette (VHS Cardiomegaly) and all four models achieved AUC of 90 % (Burti et al., 2020). Y. Yoon et al. also applied CNN to diagnose cardiomegaly, the results (accuracy, sensitivity and specificity) surpassed 95 % (Yoon et al., 2018). Here, we report a platform to detect 16 key points located on the vertebra and heart in X-ray image for dogs through a deep learning model, which can facilitate the calculation of the VHS values. These detection results can be subsequently used to assist to determine whether the examined dog’s heart is enlarged with breed-specifific or no breed specific VHS reference range (James, 2000; Lamb et al., 2001). This solution can be an alternative new route in veterinary medicine. This platform also achieves an average performance (AP) of 90.9 % and has been deployed in real clinical scenario to help doctors diagnose this disease so that both their working efficacy and efficiency were greatly improved.

Section snippets

Dataset

2670 dogs received lateral X-ray imaging, where 1875 images were used to train a deep learning model and another 399 images to validate the performance of this model. Finally, external testing dataset (396 images) were used to test the generality of this model.

16 key points are grouped into two groups: 12 key points located on the vertebra and four key points located on the heart of a dog. These images were collected in five hospitals belonging to New Ruipeng Pet Healthcare Group Co. LTD., the

Results and discussion

The detection results are shown in Fig. 5, where blue and red dots are ground truth and predicted results respectively. The corrected results after post processing are shown in Fig. 6, which shows the adjusted detection results of the images in Fig. 3 and the improvement is significant after post processing. Reference range of VHS (the approximate range is 8.5–10.7) for diagnosing the canine cardiac enlargement is different according to different dog breeds (James, 2000; Lamb et al., 2001). The

Conclusion

Current research adopted lateral X-ray image and key point detection to assist veterinarians in diagnosing canine cardiac enlargement. We applied two HRNets to detect 16 key points (12 key points on vertebrae and four on heart), and the average performance is 90.9 %. Deep learning can effectively assist doctors with canine cardiac enlargement diagnosis and veterinarians can make clinical decisions with the assistance of this system. This method can effectively avoid the subjectivity and reduce

Authors’ contributions

K.Z. designed the research; M.N.Z. and K.Z. conducted the study; Z.W.L., D.C.C., Q.R.X. and D.X.X. collected and annotated the images in the dataset; M.N.Z.was responsible for coding; M.N.Z. and K.Z. analyzed and finished experimental result; D.Y.Y., C.F.L., M.N.Z., K.Z., Q.R.X., B.L.L., D.C.C., D.X.X. co-wrote the manuscript; All authors discussed the results and commented on the manuscript.

Declaration of Competing Interest

The authors declare no competing financial interest.

Acknowledgements

No ethical committee approval was needed. Informed consent regarding personal data processing was obtained from the owners. The authors declare no competing financial interest.

References (24)

  • R. Bradley et al.

    Predicting early risk of chronic kidney disease in cats using routine clinical laboratory tests and machine learning

    J. Vet. Intern. Med.

    (2019)
  • A. Gergely et al.

    Reliability of family dogs’ sleep structure scoring based on manual and automated sleep stage identification

    Animals

    (2020)
  • Cited by (8)

    View all citing articles on Scopus
    1

    These authors contribute equally to this work.

    View full text