| Rotator cuff tear is a leading cause of shoulder pain,with an incidence rate of 22.1%,and the most common type is supraspinatus tear.Delayed diagnosis and inappropriate treatment selection for supraspinatus tear can lead to biomechanical abnormalities,decreased stability,worsening shoulder pain,and limited mobility.Providing comprehensive decision support for the prevention,diagnosis,and treatment of supraspinatus tear throughout the patient journey is crucial for optimal clinical outcomes and essential for clinical decision-making.This study provides decision support for the entire process of prevention,diagnosis,and treatment of rotator cuff tears.Focusing on risk assessment,intelligent assisted diagnosis,and postoperative prognosis prediction of rotator cuff tears,we conducted the following research based on deep learning techniques and shoulder joint MRI data.(1)Semi-automatic measurement and clinical validation of critical shoulder angle and acromion index using R-Bone Vi T.Background: Abnormal shoulder bone morphology is a major extrinsic factor for rotator cuff tears and can be used to assess tear risk,such as the critical shoulder angle and acromion index based on X-rays.However,their accuracy is greatly affected by imaging position,and three-dimensional stereoscopic evaluation remains challenging.Objective: This part aims to automatically segment and reconstruct the humerus and scapula in 3D,measure the 3D critical shoulder angle and acromion index,and validate the accuracy,reliability,and ability to assess rotator cuff tear risk.Methods: We retrospectively collected 105 shoulder MRI data as the model development dataset v1 and 369 shoulder MRI data of patients aged ≥45 years as the clinical validation dataset v1.The R-Bone Vi T segmentation network was trained on v1 to achieve 3D automatic segmentation and reconstruction of the humerus and scapula.The critical shoulder angle and acromion index were semi-automatically measured on the 3D reconstructions.Bland-Altman analysis and intraclass correlation coefficient(ICC)were used to validate the accuracy and reliability of the measurements on 30 randomly selected cases.The consistency between automatic and manual reconstructions was evaluated to validate measurement accuracy.Two attending physicians independently performed semi-automatic measurements to assess intra-and inter-observer consistency for reliability evaluation.On dataset v1,receiver operating characteristic(ROC)curve analysis determined the thresholds of critical shoulder angle and acromion index for high and low risk of rotator cuff tears based on the maximum F1 score.Results: On model development dataset v1,the automatic segmentation achieved a Dice score of 91.35±4.39 for the humerus and 79.45±5.33 for the scapula.The 95% consistency limits between automatic and manual measurements were 0.967 for both critical shoulder angle(ICC=0.828,p<0.01)and acromion index(ICC=0.866,p<0.01),demonstrating measurement accuracy.For critical shoulder angle,the intra-observer 95% consistency limit was 0.90(ICC=0.929,p<0.01),and the inter-observer 95% consistency limit was 0.90(ICC=0.948,p<0.01).For acromion index,the intra-observer 95% consistency limit was 0.933(ICC=0.926,p<0.01),and the inter-observer 95% consistency limit was 0.90(ICC=0.941,p<0.01),proving high repeatability.On clinical validation dataset v1,patients with rotator cuff tears had significantly higher critical shoulder angle(34.54±4.72 vs.30.33±4.16,p<0.01)and acromion index(0.64±0.08 vs.0.57±0.08,p<0.01)than those without tears.ROC analysis showed that the critical shoulder angle(AUC=0.778,p<0.05)performed better than the acromion index(AUC=0.746,p<0.05)in assessing tear risk.The thresholds for high tear risk were >31.24 for critical shoulder angle and >0.62 for acromion index based on the maximum F1 score.Conclusion: The R-Bone Vi T segmentation network enables fully automatic 3D reconstruction of the humerus and scapula.The semi-automatic measurement method for critical shoulder angle and acromion index based on the 3D reconstructions has high accuracy and repeatability and can assess the risk of rotator cuff tears,providing decision support for early rehabilitation intervention in shoulder pain patients.(2)Development and clinical validation of a 2D-Xception algorithm for assisted diagnosis of rotator cuff tears.Background: Compared to experienced physicians,primary care providers or clinicians with limited case experience have lower sensitivity(by 0.25)in diagnosing rotator cuff tears,leading to potential missed or misdiagnoses.Objective: This part aims to develop a 2D-Xception model for automatic diagnosis of rotator cuff tears,evaluate its diagnostic performance against clinical physicians and radiologists on different test datasets,and validate its clinical applicability.Methods: We retrospectively collected 558 shoulder MRI data as the model development dataset v2 for training the diagnostic model and 143 data as the internal test set.Additionally,69 cases of arthroscopic rotator cuff repair surgery were collected as the arthroscopic gold standard dataset for objective evaluation of model and expert performance.We trained 2D-Xception and 3D-Xception deep learning models for automatic diagnosis of rotator cuff tears.Model diagnostic performance was evaluated using sensitivity,specificity,precision,accuracy,and F1 score.We compared the model’s diagnostic performance on the internal test set and arthroscopic gold standard dataset against 8 radiologists and sports medicine physicians with varying experience levels.Subgroup analyses were conducted based on MRI field strength(1.5T and 3.0T)and tear severity to validate the model’s robustness.Results : The 2D-Xception model demonstrated the best diagnostic performance on both test datasets.On the arthroscopic gold standard dataset,the area under the ROC curve reached 0.921(0.841,1.000),and the model’s F1 score was 0.824,achieving diagnostic performance between attending sports medicine physicians(F1=0.852)and attending radiologists(F1=0.819).Subgroup analysis showed that the model had a sensitivity ranging from 0.625 to 1.000 for different tear severities on the internal test set.There was no significant performance difference between 1.5T and 3.0T MRI data(P>0.05).Conclusion : The 2D-Xception model achieves efficient and accurate automatic diagnosis of rotator cuff tears,reaching a diagnostic performance comparable to attending physicians.It can provide diagnostic decision support for primary care facilities and less experienced clinicians.(3)Development and clinical validation of a quantitative assessment method for supraspinatus muscle atrophy using R-Muscle UNet.Background : The degree of supraspinatus muscle atrophy is a major indicator for evaluating the risk of re-tear after rotator cuff repair surgery.However,current assessments are mostly qualitative and subjective,influenced by single-plane limitations and clinical expertise,lacking quantitative studies.Objective: This part aims to automatically segment the supraspinatus muscle and fossa,quantify the relative volume of the supraspinatus muscle to assess its atrophy,analyze the risk of postoperative re-tear,and validate its clinical applicability.Methods: We retrospectively collected 511 shoulder MRI data as the model development dataset v3 and 2,805 shoulder MRI data as the clinical validation dataset v2.Additionally,27 cases with complete preoperative MRI and postoperative follow-up after arthroscopic rotator cuff repair were included as the postoperative prognosis dataset.The R-Muscle UNet segmentation network was trained on v3 to achieve automatic selection of sagittal Y-plane slices and segmentation of the supraspinatus muscle and fossa.The relative volume of the supraspinatus muscle was calculated as the ratio of its volume to the fossa volume to assess muscle atrophy.Bland-Altman analysis was used to evaluate the consistency between automatic and manual segmentation-based volume ratio measurements for validating measurement accuracy.On v2,univariate and multivariate linear regressions were performed to assess the relationships between supraspinatus muscle atrophy,age,and tear severity.On the postoperative prognosis dataset,ROC curve analysis determined the threshold of relative supraspinatus muscle volume for high and low risk of postoperative re-tear based on the maximum F1 score.Results:The R-Muscle UNet network achieved a slice selection accuracy of 0.9147,with Dice scores of 93.64±0.72 for supraspinatus muscle segmentation and 93.73±0.70 for supraspinatus fossa segmentation,demonstrating excellent segmentation performance.The 95% consistency limit between the automatic and manual measurements of the relative supraspinatus muscle volume reached 0.945.On the clinical validation dataset v2,univariate linear regression analysis showed that age and different tear severities significantly influenced the relative supraspinatus muscle volume(all p<0.01),indicating that age and the severity of supraspinatus muscle tears are contributing factors to muscle atrophy.As age increases and tear severity worsens,supraspinatus muscle atrophy becomes more severe.Multivariate linear regression analysis revealed that age,grade 2 full-thickness tears,and grade 3 full-thickness tears(all p<0.01)were independent factors influencing supraspinatus muscle atrophy.On the postoperative prognosis dataset,ROC analysis demonstrated that for postoperative re-tear,the area under the curve(AUC)for the relative supraspinatus muscle volume ratio reached 0.75(0.529,0.971).With the maximum F1 score,the volume ratio threshold was 0.576,indicating a high risk of postoperative re-tear for patients with a volume ratio less than 0.576.Conclusion: The R-Muscle UNet segmentation network enabled quantitative assessment of supraspinatus muscle atrophy,demonstrating that increasing age and worsening tear severity exacerbate supraspinatus muscle atrophy.The relative supraspinatus muscle volume can effectively evaluate the risk of postoperative re-tear in patients with rotator cuff tears,providing decision support for treatment planning. |
