| Summary: | The purpose of this study was to evaluate kinematic analysis repeatability by deep learning
approach in countermovement jump. Seventy athletes (39 women, 31 men) performed two
maximal countermovement jumps in either one session or two separate sessions (jumps
separated by two-weeks). The jumps were filmed from lateral and frontal point of view. Video
data from 50 athletes were selected randomly to be used for training the deep learning model
with DeepLabCut. A total of 10 images were used from every athlete from this training set,
meaning that a total of 500 images were used to create the model for frontal view and side view
(sagittal) videos. The performance of this model was then evaluated by applying it on 11 withinday measurements and 9 between-day measurements again for both frontal and sagittal videos.
For frontal view videos, the marker locations were labelled for both sides of the body to
shoulder (acromion), hip joint (greater trochanter), knee joint (mid-point of patella) ankle joint
(mid-point between malleoli) and toes (head of shoe). The marker locations of shoulder
(acromion), hip joint (greater trochanter), knee joint (lateral femoral condyle), ankle joint
(lateral malleolus) and toes (head of shoe) were manually labelled for sagittal test images. For
the sagittal videos, hip, knee and ankle joint angles were calculated by using atan2 function in
Matlab, and for the frontal view videos, the same was done for the knee and ankle angles. To
compensate for misplaced or missing markers, raw data was filtered with a median filter and
subsequently with Butterworth 4th order low-pass filter. After filtering, data was further
processed with Matlab by first aligning the curve data of consecutive (trial 1 and trial 2) jumps.
Then data was cropped according to the movement of knee joint from sagittal plane: start of
cropping was selected as the point where there was a 5-degree joint angle change from the
initial standing position, and the end point was selected as the same calculated value after
landing the countermovement jump. Test-retest values were calculated with intraclass
correlation coefficients (ICC) for subjects in the evaluation set. The ICC model used for testretest was single measurement two-way mixed effects with absolute agreement. High mean ICC
values were observed for sagittal within-day joint angles (0.95 ± 0.04 for hip joint, 0.96 ± 0.03
for knee joint and 0.95 ± 0.05 for ankle joint). Similar values were found for mean betweenday measurements (0.95 ± 0.03 for hip joint, 0.95 ± 0.07 for knee joint and 0.89 ± 0.08 for ankle
joint). On the contrary, correlations of joint angle values for frontal plane varied substantially
more: For within-day measurements, mean ICC values revealed poor test-retest reliability for
right knee angle (ICC = 0.43 ± 0.31), and moderate test-retest reliability for left knee (ICC =
0.68 ± 0.23), right ankle (ICC = 0.62 ± 0.22) and left ankle (ICC = 0.53 ± 0.29) angles. Mean
between-day ICC values demonstrated good (ICC = 0.75 ± 0.10) test-retest reliability for right
knee angle, moderate test-retest reliability for left ankle angle (0.53 ± 0.17), and poor test-retest
reliability for left knee (ICC = 0.49 ± 0.27) and right ankle (ICC = 0.34 ± 0.26) angles. These
results imply that deep learning approach provides very repeatable measurements for sagittal
joint angles in countermovement jump, but not as such for frontal plane kinematics. Hence deep
learning approach provides an affordable and easy-to-access method to perform repeated
measurements for 2-D motion analysis of countermovement jump and possibly other sports
movements filmed from sagittal plane. Further studies on repeatability and the validation of
deep learning-based systems are required to prove their accuracy and to provide reliable data
for practitioners.
|
|---|