Madison Wheeler
Pronouns: she/her
Research Mentor(s): Riann Smith
Research Mentor School/College/Department: School of Kinesiology / Kinesiology
Program: UROPF
Session: Session 5 (2:40pm – 3:30pm)
Authors: Madison Wheeler, Beyza Tayfur, Alexa Johnson, Riann Smith
Abstract
Background The cross-sectional area of muscles from ultrasound imagery is generally evaluated manually to produce numerical data, a process that is laborious and time-consuming. A new artificial intelligence program (DeepACSA) can estimate cross-sectional area quickly and accurately with little human intervention necessary.1 However, this program was created and validated using healthy muscles and it remains unclear if it would be effective in patients whose muscles are affected by injury, such as those who have sustained an anterior cruciate ligament (ACL) injury. Therefore, the purpose of this study is to evaluate the accuracy of DeepACSA in assessing cross-sectional area of the quadriceps muscles in people who have sustained an ACL injury and subsequently underwent ACL reconstruction. Methods The participants used in this study are part of a larger double-blind randomized control trial with a goal of improving muscle strength and function after ACL reconstruction. Muscle ultrasound images of the rectus femoris pre-reconstruction (n=50 images) and two months post- reconstruction (n=50 images) were collected via a GE ultrasound. All images were entered into DeepACSA to calculate the rectus femoris cross-sectional area (in cm2). The results from DeepACSA were visually checked to ensure accuracy of the muscle borders that predicted cross-sectional area and only accurate images were used during the cross-examination. Then, using ImageJ Script2, all images were manually traced to calculate cross-sectional area. Finally, manual results were compared with DeepACSA results using the intraclass correlation coefficient (ICC2,1 , intra-rater reliability, two-way random effects, single rater, absolute agreement).3 Results Rectus femoris cross-sectional area was predicted accurately by DeepACSA in 63% of pre-reconstruction images and 60% of the two months post-reconstruction images. The ICC2,1 for pre-reconstruction was .937 (95% confidence interval: .863 to .972) and for two months post-reconstruction was .967 (95% confidence interval: .893 to .987). These values suggest good to excellent agreement between DeepACSA and manual analysis. Conclusion DeepACSA accurately predicted borders to calculate cross-sectional area in ~60% of images. Although the program does not accurately identify muscle borders in all images, those with correctly identified borders had similar cross-sectional areas to manual analysis, as suggested by the high ICC values at both time points. Further work should be done to improve the accuracy with which DeepACSA identifies muscle borers in injured muscles. With additional improvements DeepACSA could minimize the need for manual analysis of ultrasound images in research studies, thus improving data processing efficiency and reducing error between raters. 1. Ritsche, Wirth, Cronin, et al. DeepACSA: Automatic Segmentation of Cross-sectional Area in Ultrasound Images of Lower Limb Muscles Using Deep Learning. Med Sci Sports Exerc, 2022; 54(12): 2188-2195. 2. Schneider, Rasband, Eliceiri. NIH Image to ImageJ: 25 years of image analysis. Nat Methods, 2012; 9: 671-675. 3. Koo, Li. A Guideline of Selecting and Reporting Intraclass Correlation Coefficients for Reliability Research. J Chiropr Med, 2016; 15(2): 155-163.
