Skip to main content
2024 BMES Annual Meeting Main banner
Icon Legend
This session is not in your schedule.
This session is in your schedule. Click again to remove it.

High School Student with a Traditionally Underrepresented Background

HS Poster

Poster 50 - Deep learning-based algorithm for automatic classification of knee osteoarthritis severity in pre-processing-free radiographs using staged ROI prediction technique

Saturday, October 26, 2024
10:00 AM - 12:00 PM EST
Location: Exhibit Hall E, F & G

    Presenting Author(s)

  • Kyungwon Shin (she/her/hers)

    High School Student
    Chadwick International School
    Yeonsu-Gu, Inch'on-jikhalsi, Republic of Korea

    • Co-Author(s)

  • Min-Kyu Lee, AS (he/him/his)

    AI Developer
    Walk101, Co.,Ltd., Republic of Korea

    • Last Author(s)

  • Dong-Sik Chae, MD–PhD (he/him/his)

    Professor, Department of Orthopedic Surgery, International St. Mary’s Hospital, Catholic Kwandong Un
    Department of Orthopedic Surgery, International St. Mary’s Hospital, Catholic Kwandong University College of Medicine
    Incheon, Inch'on-jikhalsi, Republic of Korea


Introduction: Knee osteoarthritis (OA) is a frequent form of osteoarthritis, a degenerative joint disease that breaks down tissue joints over time and causes pain when moving the knees [1]. OA typically affects individuals over the age of 55. With an aging population, OA is expected to become more and more prevalent [2]. By 2050, there may be approximately 642 million people worldwide with OA, a 74.9% increase from 2020 [3]. Patients with OA are often limited in physical activity and experience decreased quality of life [2]. Therefore, it is crucial to detect OA at an early stage to control its progression. The most common form of detection involves a clinician using the Kellgren-Lawrence grade [4] to classify the grade of OA from a radiograph. However, there are several challenges, including the limited availability of radiologists and orthopedic specialists and the fact that the results depend on the clinician’s subjective assessment. Therefore, using artificial intelligence models for automatic, consistent, and accurate classification yields promising outcomes that benefit both patients and clinicians. Previous research has already concluded that these models demonstrate significant levels of accuracy, with the highest being 98.36% [4]. However, the accuracy is limited to an ideal situation where the input image is preprocessed to display the active imaging area only. Therefore, this study was conducted to develop an algorithm that can accurately assess the KL grade from the original radiograph using the ROI (region of interest) prediction technique, making it applicable in real-life clinical practice.

Materials and

Methods: The raw dataset consisted of 650 knee X-rays taken in Rosenberg view sourced from Catholic Kwandong University International St. Mary's Hospital. The images were anonymized, and the hospital's Institutional Review Board approved the study. The proposed algorithm has two roles: preparing the datasets and assessing the KL grades. Assessing the KL grade for each knee joint involves identifying two KL grades, each for the joint’s medial and lateral compartments, and then adopting the highest grade observed. If each compartment is not separated, the model will look at two different grades simultaneously. We hypothesized that this may hinder the model's performance and, therefore, created an algorithm to separate the compartments. The algorithm uses YOLO, and as seen in Figure 1, two stages occur simultaneously: the first is distinguishing and cropping the left and right knee joints (first ROI area), and the second is distinguishing and cropping the medial and lateral compartments of the knee joints (second ROI area). In the end, a total of 2500 image files were created, each with a distinct compartment. An orthopedics specialist classified all compartments according to the KL grade. Then, we created two datasets: a whole knee joint dataset composed of both compartments and a compartment-separate dataset. We trained three SOTA models (MobileNet, NASNetLarge, and ResNet152V2) that were concluded to have a validation accuracy of over 90% [4] with the two datasets to classify the KL grade. A preliminary pilot study was conducted with a small sample size of 80 image files and with MobileNet.

Results, Conclusions, and Discussions: The results from the pilot study yielded an overall high success rate, as seen in Figure 2. The data preparation step of detecting and cropping the knee joints and compartments was accurate, with YOLO yielding a 97.1% accuracy. MobileNet achieved high levels of precision with a validation accuracy of 90.9%. Considering the small dataset used to train the model in the pilot study, we expect to obtain higher accuracy levels when the model is trained with the complete dataset.

Interpreting how DL algorithms work benefits both the fields of computer science and healthcare, as they enable the development of improved AI algorithms and thus advance their utility in healthcare systems. Significant work has already been done with AI in healthcare, and the quality of treatment has improved. Organizations such as Google and IBM have developed DL models that make predictions around hospitalized patients and support managing patient data and results [5]. As we interpret and understand how successful models work, more and more AI and DL tools will likely be developed to address particular needs within healthcare, including orthopedic diagnostics.

The present study paves the way for using artificial intelligence to develop an automatic, consistent, and accurate OA classification tool that is applicable in real-life clinical practice. The methods presented in existing papers show sufficient accuracy in predicting KL grade. However, there is a limitation that X-rays taken in actual clinical settings must be preprocessed through measures such as cropping and resizing before being input into the prediction program. Therefore, we have sought to develop a novel system that performs simultaneous preprocessing and prediction without prior processing of X-rays. Such systems will be able to aid clinician's decision-making in real-life clinical practice.

One limitation of the study is that the datasets we trained the models with are from a single institution. We recommend that future studies will incorporate datasets from multiple institutions to assess the generalizability of the proposed method on a larger scale.

Acknowledgements (Optional): Corresponding author
Name: Dong-Sik Chae, M.D., Ph.D.
Affiliation: Department of Orthopedic Surgery, International St. Mary’s Hospital, Catholic Kwandong University College of Medicine, Incheon, Republic of Korea
Address: 22711, 25, Simgok-ro 100beon-gil, Seo-gu, Incheon, Republic of Korea
Telephone: +82-32-290-3878
Fax number: +82-32-290-3879
E-mail: drchaeos@gmail.com
ORCID ID: 0000-0002-7936-088X

Co-author
Name: Min-Kyu Lee, AI Developer
Affiliation: Walk101, Co.,Ltd., Incheon, Republic of Korea

References (Optional): [1] National Institute of Arthritis and Musculoskeletal and Skin Diseases. (2019, August 2). Osteoarthritis. National Institute of Arthritis and Musculoskeletal and Skin Diseases. https://www.niams.nih.gov/health-topics/osteoarthritis

[2] World Health Organization. (2023, July 14). Osteoarthritis. www.who.int; World Health Organization. https://www.who.int/news-room/fact-sheets/detail/osteoarthritis

[3] Steinmetz JD, et al. The Lancet Rheumatol. 2023; 5(9): E508-E522

[4] Vijaya Kishore V, Kalpana V, Kumar GH. E-Prime. 2023; 5: 100266

[5] Tiwari A, Poduval M, Bagaria V. World J Orthop. 2022; 13(6): 603–614