Deep learning for the rapid automatic quantification and characterization of rotator cuff muscle degeneration from shoulder CT datasets.

OBJECTIVES: This study aimed at developing a convolutional neural network (CNN) able to automatically quantify and characterize the level of degeneration of rotator cuff (RC) muscles from shoulder CT images including muscle atrophy and fatty infiltration. METHODS: One hundred three shoulder CT scans from 95 patients with primary glenohumeral osteoarthritis undergoing anatomical total shoulder arthroplasty were retrospectively retrieved. Three independent radiologists manually segmented the premorbid boundaries of all four RC muscles on standardized sagittal-oblique CT sections. This premorbid muscle segmentation was further automatically predicted using a CNN. Automatically predicted premorbid segmentations were then used to quantify the ratio of muscle atrophy, fatty infiltration, secondary bone formation, and overall muscle degeneration. These muscle parameters were compared with measures obtained manually by human raters. RESULTS: Average Dice similarity coefficients for muscle segmentations obtained automatically with the CNN (88% ± 9%) and manually by human raters (89% ± 6%) were comparable. No significant differences were observed for the subscapularis, supraspinatus, and teres minor muscles (p > 0.120), whereas Dice coefficients of the automatic segmentation were significantly higher for the infraspinatus (p < 0.012). The automatic approach was able to provide good-very good estimates of muscle atrophy (R2 = 0.87), fatty infiltration (R2 = 0.91), and overall muscle degeneration (R2 = 0.91). However, CNN-derived segmentations showed a higher variability in quantifying secondary bone formation (R2 = 0.61) than human raters (R2 = 0.87). CONCLUSIONS: Deep learning provides a rapid and reliable automatic quantification of RC muscle atrophy, fatty infiltration, and overall muscle degeneration directly from preoperative shoulder CT scans of osteoarthritic patients, with an accuracy comparable with that of human raters. KEY POINTS: • Deep learning can not only segment RC muscles currently available in CT images but also learn their pre-existing locations and shapes from invariant anatomical structures visible on CT sections. • Our automatic method is able to provide a rapid and reliable quantification of RC muscle atrophy and fatty infiltration from conventional shoulder CT scans. • The accuracy of our automatic quantitative technique is comparable with that of human raters.

Optimization of surgical parameters based on patient-specific models: Application to arcuate keratotomy.

PURPOSE: To determine surgical parameters for arcuate keratotomy by simulating the intervention with a patient-specific model. SETTING: University Eye Clinic Salzburg, Paracelsus Medical University, Austria, and Institute for Surgical Technology and Biomechanics, University of Bern, Switzerland. DESIGN: Computational modeling study. METHODS: A new approach to plan arcuate keratotomy based on personalized finite element simulations was developed. Using this numeric tool, an optimization algorithm was implemented to determine the incision parameters that best met the surgeon's requirements while preserving the orientation of the astigmatism. Virtual surgeries were performed on patients to compare the performance of the simulation-based approach with results based on the Lindstrom and Donnenfeld nomograms and with intrastromal interventions. RESULTS: Retrospective data on 28 patients showed that personalized simulation reproduced the surgically induced change in astigmatism (Pearson correlation = 0.8). Patient-specific simulation was used to examine strategies for arcuate interventions on 621 corneal topographies. The Lindstrom nomogram resulted in low postoperative astigmatism (mean 0.03 diopter [D] ± 0.3 [SD]) but frequent overcorrections (20%). The Donnenfeld nomogram and intrastromal incisions resulted in a small amount of overcorrection (1.5%) but a wider spread in astigmatism (mean 0.63 ± 0.35 D and 0.48 ± 0.50 D, respectively). In contrast, the new numeric parameter optimization approach led to postoperative astigmatism values (mean 0.40 ± 0.08 D, 0.20 ± 0.08 D, and 0.04 ± 0.13 D) that closely matched the target astigmatism (0.40 D, 0.20 D, and 0.00 D), respectively, while keeping the number of overcorrections low (<1.5%). CONCLUSION: Using numeric modeling to optimize surgical parameters for arcuate keratotomy led to more reliable postoperative astigmatism, limiting the risk for overcorrection.