Development of artificial intelligence-based clinical decision support system for diagnosis of meniscal injury using magnetic resonance images. (April 2023)

• Record Type: Journal Article
• Authors: Chou, Yi-Ting; Lin, Ching-Ting; Chang, Ting-An; Wu, Ya-Lun; Yu, Cheng-En; Ho, Tsung-Yu; Chen, Hui-Yi; Hsu, Kai-Cheng; Kuang-Sheng Lee, Oscar
• Is Part Of: Biomedical signal processing and control. Volume 82(2023)
• Publication Date: 2023-04
• Subjects: Artificial Intelligence -- Magnetic Resonance Imaging -- Meniscal tear diagnosis -- Arthroscopy -- Knee; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104523
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: This article aimed to utilize magnetic resonance images and arthroscopic findings as a dataset to develop an artificial intelligence model that can help detect meniscal lesions, and thus aid in the diagnosis of meniscus injuries. Our dataset comprised data from 811 knee magnetic resonance imaging studies. magnetic resonance images were labeled and annotated by two orthopedic surgeons. First, Scaled-YOLOv4 was used to detect the position of the meniscus. Second, the EfficientNet-B7 model structure was used to detect meniscal tears. Through the artificial intelligence detection system, clinicians can obtain a structured report of meniscus rupture, which enables physicians to interpret images more quickly and saves time for physicians to read images. Abstract: Background: Magnetic resonance imaging (MRI) examinations are often necessary for the diagnosis of meniscal injuries. Due to the quantity and detail of the images in an MRI examination, a careful reading of MR images is time-consuming even for experienced physicians. The use of artificial intelligence (AI) models for interpreting MR images may address this problem. Objective: This study aimed to utilize MR images and arthroscopic findings as a dataset to develop an AI model that can help detect meniscal lesions, and thus aid in the diagnosis of meniscus injuries. Method: Our dataset comprised data from 811 knee MRI studies. MR images were labeled and annotated by two orthopedic surgeons. The training pipeline had two parts. First, Scaled-YOLOv4 was used to detect the position of the meniscus. Second, the EfficientNet-B7 model structure was used to detect meniscal tears. A brief report demonstrating the probability and the position of a normal or torn meniscus was then generated by the AI-based clinical decision support system. Result: The Scaled-YOLOv4 model had areas under the curve (AUCs) of 0.948 and 0.963 in the sagittal and coronal views, respectively. The EfficientNet-B7 model yielded AUCs of 0.984 and 0.972 in the sagittal and coronal views, respectively. High AUCs were achieved both in meniscus localization and meniscal tear detection. Discussion: In this study, we utilized arthroscopic findings, obtained using the optimal meniscal tear diagnosis method, as our ground truth label. Nevertheless, this approach did not provide better model performance than that based on MRI clinical reports. Conclusions: The contributions of this paper include the following points. First, an AI-based clinical decision support system was established to detect meniscal injury based on a deep learning model algorithm. High sensitivity and specificity were achieved by the model system. The model may have substantially facilitated the diagnosis of meniscal injuries from MR images. Second, through the AI detection system, clinicians can obtain a structured report of meniscus rupture, which enables physicians to interpret images more quickly and saves time for physicians to read images.