Fully automatic algorithm for detecting and tracking anatomical shoulder landmarks on fluoroscopy images with artificial intelligence

OBJECTIVE

Patients with rotator cuff tears present often with glenohumeral joint instability. Assessing anatomic angles and shoulder kinematics from fluoroscopy requires labelling of specific landmarks in each image. This study aimed to develop an artificial intelligence model for automatic landmark detection from fluoroscopic images for motion tracking of the scapula and humeral head.

MATERIALS AND METHODS

Fluoroscopic images were acquired for both shoulders of 25 participants (N = 12 patients with unilateral rotator cuff tear, 6 men, mean (standard deviation) age: 63.7 ± 9.7 years; 13 asymptomatic subjects, 7 men, 58.2 ± 8.9 years) during a 30° arm abduction and adduction movement in the scapular plane with and without handheld weights of 2 and 4 kg. A 3D full-resolution convolutional neural network (nnU-Net) was trained to automatically locate five landmarks (glenohumeral joint centre, humeral shaft, inferior and superior edges of the glenoid and most lateral point of the acromion) and a calibration sphere.

RESULTS

The nnU-Net was trained with ground-truth data from 6021 fluoroscopic images of 40 shoulders and tested with 1925 fluoroscopic images of 10 shoulders. The automatic landmark detection algorithm achieved an accuracy above inter-rater variability and slightly below intra-rater variability. All landmarks and the calibration sphere were located within 1.5 mm, except the humeral landmark within 9.6 mm, but differences in abduction angles were within 1°.

CONCLUSION

The proposed algorithm detects the desired landmarks on fluoroscopic images with sufficient accuracy and can therefore be applied to automatically assess shoulder motion, scapular rotation or glenohumeral translation in the scapular plane.

CLINICAL RELEVANCE STATEMENT

This nnU-net algorithm facilitates efficient and objective identification and tracking of anatomical landmarks on fluoroscopic images necessary for measuring clinically relevant anatomical configuration (e.g. critical shoulder angle) and enables investigation of dynamic glenohumeral joint stability in pathological shoulders.

KEY POINTS

• Anatomical configuration and glenohumeral joint stability are often a concern after rotator cuff tears. • Artificial intelligence applied to fluoroscopic images helps to identify and track anatomical landmarks during dynamic movements. • The developed automatic landmark detection algorithm optimised the labelling procedures and is suitable for clinical application.

Load-induced increase in muscle activity during 30° abduction in patients with rotator cuff tears and control subjects

BACKGROUND

Rotator cuff muscles stabilise the glenohumeral joint and contribute to the initial abduction phase with other shoulder muscles. This study aimed to determine if the load-induced increase in shoulder muscle activity during a 30° abduction test is influenced by asymptomatic or symptomatic rotator cuff pathologies.

MATERIALS AND METHODS

Twenty-five patients with unilateral rotator cuff tears (age, 64.3 ± 10.2 years), 25 older control subjects (55.4 ± 8.2 years) and 25 younger control subjects (26.1 ± 2.3 years) participated in this study. Participants performed a bilateral 30° arm abduction and adduction movement in the scapular plane with handheld weights (0-4 kg). Activity of the deltoid, infraspinatus, biceps brachii, pectoralis major, latissimus dorsi and upper trapezius muscles was analysed at maximum abduction angle after normalisation to maximum voluntary contraction. Shoulders were classified into rotator cuff tendinopathy, asymptomatic and symptomatic rotator cuff tears, and healthy based on magnetic resonance images. A linear mixed model (loads, shoulder types) with random effects (shoulder identification) was applied to the log-transformed muscle activities.

RESULTS

Muscle activity increased with increasing load in all muscles and shoulder types (P < 0.001), and 1-kg increments in additional weights were significant (P < 0.001). Significant effects of rotator cuff pathologies were found for all muscles analysed (P < 0.05). In all muscles, activity was at least 20% higher in symptomatic rotator cuff tears than in healthy shoulders (P < 0.001). Symptomatic rotator cuff tears showed 20-32% higher posterior deltoid (P < 0.05) and 19-25% higher pectoralis major (P < 0.01) activity when compared with asymptomatic tears.

CONCLUSIONS

Rotator cuff pathologies are associated with greater relative activity of shoulder muscles, even with low levels of additional load. Therefore, the inclusion of loaded shoulder tests in the diagnosis and rehabilitation of rotator cuff pathologies can provide important insight into the functional status of shoulders and can be used to guide treatment decisions.

LEVEL OF EVIDENCE

Level 2.

TRIAL REGISTRATION

Ethical approval was obtained from the regional ethics committee (Ethics Committee Northwest Switzerland EKNZ 2021-00182), and the study was registered at clinicaltrials.gov on 29 March 2021 (trial registration number NCT04819724, https://clinicaltrials.gov/ct2/show/NCT04819724 ).