Introduction: This study developed an open access workflow for semi-automated three dimensional (3D) alpha angle measurement to allow researchers and clinicians to evaluate the 3D degree of cam morphology in a fast and objective manner, regardless of their financial resources. Method: In the semi-automated workflow geometrical approaches were applied to eight manually segmented 3D models of the proximal femur (four left and four right). These 3D models were created from Zero Echo Time (ZTE) Magnetic Resonance Imaging (MRI) images of four randomly selected individuals from the Feyenoord cohort, thus without radiation exposure. Important landmarks were localized on the 3D models with subsequent radial reformatting and alpha angle computation in 12 radial sections. Novel strategies included radial section-specific spherical least square fitting to approximate the radius of the femoral head, outlier detection to handle surface irregularities and model specific thresholds to account for the asphericity of the femoral head. For six clockwise positions (12:00, 1:00, 2:00, 3:00, 6:00, and 9:00 o’clock), both qualitative and quantitative validation were conducted by comparing the semi-automated measurements to manual alpha angle measurements performed by an experienced musculoskeletal (MSK) radiologist. To evaluate the inter-observer reliability, the mean absolute difference (MAD) and intra-class correlation (ICC) were calculated for various automatic approaches, depending on which novel strategies were incorporated. Additionally, the radii (mm) and the measurement time (minutes) were compared between the manual and semi-automated approach. Results: Reliability assessment showed poor agreement with ICCs ranging from -0.02 to 0.40, depending on which automatic approach was used. The highest ICC of 0.40 [95%CI: -0.51 - 0.89] was achieved using a radial section-specific radius without alteration of the spherical threshold or application of outlier detection. MADs ranged from 20°- 23°. Statistically significant differences (p-value <0.003) between manually drawn circular radii and automatically fitted radial section-specific spherical radii were found for 50% of the 3D hip models. All manually drawn circular radii were smaller than the radii of the automated fitted spheres. Variation in radii between and within radial sections were qualitatively and quantitatively demonstrated. Lastly, computation of the 12 clockwise alpha angles was performed by the workflow within 26 seconds, indicating a time difference of approximately 15 minutes compared to manual measurements. Conclusion: The proposed semi-automated workflow is capable of computing 12 alpha angles across radial sections surrounding the femoral head in a fast and objective manner. The performance of the workflow could be improved by refining landmark localization, enhancing the quality and segmentation of ZTE MRI scans, enlarging the validation dataset and analysis of prior workflow error propagation. This study established that the femoral head is not perfectly spherical. This affects the accuracy of alpha angle measurements as it is strongly sensitive to radius estimation and therefore questions the reliability of the alpha angle as diagnostic measure of cam morphology. Future research should explore the potential of complementary alternative imaging parameters such as the Head-Neck-Offset Ratio (HNOR) and Triangular Index (TI), as well as the use of statistical shape models that incorporate the imperfect spherical morphology of the femoral head.

Background: Advancements in the field of artificial intelligence have lead to the incorporation of automated algorithms in the analysis of medical images and data. Deep learning algorithms have been applied in muscu- loskeletal research to improve the understanding of osteoarthritis and to assist in disease detection and prognosis. The majority of the developed methods examine and process X-ray images and clinical data (age, gender etc.), with a small minority using MRI as inputs. Objective: The current master thesis project aims to investigate the influence of MRI scans on the early detection of knee osteoarthritis through the use of deep learning architectures, and to develop a semi-automatic method for knee region of interest extraction for creating the MRI input of detection algorithms. Methods: The MRI scans used in this project were acquired from the publicly available database of the Os- teoarthritis Initiative. In total 593 dual echo steady state and intermediate-weighted turbo spin-echo sequences were included. The extraction of the knee joint included several processing steps. Initially, a U-Net model was trained on 507 annotated dual echo steady state MRIs for the segmentation of bone and cartilage tissue, which was followed by the registration of the output masks to intermediate-weighted turbo spin-echo sequences in order to create the joint labels for the desired MRI scans. Final step for the region of interest construction included the search of bone coordinates and the creation of the knee joint region of interest. The detection of early osteoarthri- tis progression from knee MRI scans was tested through three different deep learning architectures, a residual network (ResNet), a densely connected convolutional network (DenseNet) and a convolutional variational au- toencoder (CVAE). Furthermore, the probability output of the ResNet and DenseNet as well as the feature vector of the CVAE were coupled with clinical data (age, gender, bone mass index) and used as input to a Logistic Regression Classifier, in order to investigate the influence of osteoarthritis related features to the detection task. The U-Net segmentation method was evaluated using Dice similarity coefficient and Intersection of Union while the detection algorithms using the area under the receiver’s characteristic curve (AUC) and the precision-recall curve (PR-AUC) metrics, with two different input data configurations, only MRI and a combination of MRI and clinical data. Results: The U-Net algorithm for bone and cartilage segmentation showed adequate results, since Dice simi- larity coefficient and Intersection of Union reached mean values higher than 0.99 and 0.88. Regarding the early detection of knee osteoarthritis incidence, ResNet and DenseNet showed similar results, with both methods hav- ing an AUC value ranging from 0.5033 to 0.6269, when only MRI scans were examined. In the case of MRI and clinical data combination, the more complicated deep learning architecture (DenseNet) achieved the highest AUC at 0.6556. The best performing model was CVAE with the largest number of latent space features (1000) achieving an AUC of 0.6699 when combined with clinical data and an AUC of 0.6689 when used alone as input to the logistic regression classifier. All three deep learning algorithms yielded higher performance metrics when clinical data where combined with models’ outputs. Conclusion: The tested deep learning algorithms showed a potential in the challenging task of early detection of knee osteoarthritis through MRI scans, even though they did not reach the same level of performance metrics. The region-of-interest creation had promising results for the implementation of U-Net method for bone tissue labelling.