This study presents the first automated workflow to analyze peri-articular bone structures and microarchitecture in the knee using high-resolution peripheral quantitative CT (HR-pQCT) scans. The framework was trained to segment 2,598 scans of the radius and tibia, which were further fine-tuned on 131 knee images, combining transfer learning and atlas-based registration to determine regions of interest. Out of all the segmentation models, the UNet++ was observed to have the most accurate segmentation of the subchondral bone plate, while the diffeomorphic demons algorithm for atlas-based registration resulted in the highest Dice similarity coefficient (DSCs) from 0.81 to 0.84. The model was validated against 128 knee scans, demonstrating strong correlations between the predicted and reference measurements (R^2 = 0.86 to 0.99) though moderate bias was present in the subchondral bone plate density and thickness measurements. When validated against another triple-repeat-measures dataset with only tibia images, the model demonstrated excellent precision and reproducibility, with the average short-term coefficient of variation (RMS%CV) ranging from 1.0% to 2.9%. However, limitations were noted, especially the subjectivity of the reference segmentations used to train the model caused by inter-observer variation in the semi-automated segmentation of the knee. Nonetheless, this workflow demonstrated unbiased, automatic, and precise measurements of the knee microarchitecture compared to semi-automated methods that were both labor-intensive, due to the sheer size of a knee HR-pQCT image, and the inter-operator variability that led to inaccuracies and bias. Importantly, the authors have also made their model, code, and atlases publicly available for researchers to utilize, accelerating the rate of progress for HR-pQCT research everywhere.