Automated Joint Space Detection Improves Bone Segmentation Accuracy

Recent high throughput image processing algorithms have provided numerous gold standard bone segmentations of complex joints with the potential to train deep learning models. The current challenge is the limited ability to utilize prior segmentation strategies in future experiments given structure-specific algorithms and significant manual corrective processes. We address the need for reproducible automated image analysis that enables quantitative assessment of bone health and osteoarthritis progression while minimizing manual labor and viability.

To begin, view the micro computed tomography dataset to observe that bone and soft tissue are clearly distinguishable. Inspect regions where adjacent bones are closely packed to identify areas where limited image resolution and minimal gaps make their boundaries difficult to detect. Note the baseline where the supervised watershed-based segmentation method produces an overall accuracy of approximately 80%To improve segmentation accuracy, apply an updated workflow that combines a pre-trained deep learning joint space prediction model with a pre-built image processing and watershed-based bone separation recipe in the Amira software environment.

Load the micro computed tomography dataset into Amira. Connect the deep learning prediction module. Select the trained model and press Apply to generate an initial joint space segmentation.

Then load the recipe in the image recipe player module. Connect the micro computed tomography dataset and the preliminary joint segmentation, and apply the recipe to produce the final segmentation with fully individualized bones. Finally, evaluate the resulting output by adjusting the color map to labels 256 and inspecting orthogonal slices and volume renderings to confirm successful separation of individual bones throughout the hindpaw.

To inspect or edit the image processing recipe, open the image recipe workroom in Amira and load the bone separation recipe. Connect the required input data and adjust parameters in the Properties window to customize the workflow for the current dataset. To review the workflow structure, sequentially visualize the intermediate outputs generated by the recipe, then apply a mathematical morphology black top-hat method implemented through closing, arithmetic subtraction and thresholding, and use this step to emphasize voxels that are darker than their surroundings and correspond to joint regions.

To improve specificity, apply the Structure Enhancement filter to enhance dark, thin planar structures corresponding to joint spaces. Further reinforce joint continuity by applying the Membrane Enhancement filter, which incorporates a tensor voting stage. Then, integrate the joint mask generated by a convolutional neural network trained to identify periarticular negative space with the results from the previous image processing filters to generate a final joint mask.

To complete bone separation, subtract the final joint mask from the bone mask and treat each remaining connected component as a marker for an individual bone. Apply the watershed algorithm to grow each marker within the original bone mask, reconstructing individual bones, and placing boundaries at optimal positions based on image intensities. To prepare training data for the deep learning model, select 20 wild-type datasets representing 40 hindpaws with balanced sex distribution and ages ranging from 2 to 6 months.

Extract six sub-volumes from each dataset with three sub-volumes per hindpaw. Use sub volumes measuring 200 x 200 x 200 voxels to cover ankle regions, digits, and representative background areas. Configure a three-dimensional U-Net architecture with a ResNet-18 backbone.

Optimize the model using the Adam algorithm with an initial learning rate of 0.0001 and randomly initialized weights. Train the model using the dice loss function and monitor validation performance using the intersection over union metric. Using deep learning to identify joint spaces in micro-CT datasets enabled clear and consistent separation of individual hindpaw bones.

By focusing on the periarticular negative space, the model accurately delineated bone boundaries even in densely packed regions. Compared with the semi-automated workflow, which achieved about 80%accuracy, the deep learning approach delivered a marked improvement for wild-type mice, reaching 98-99%accuracy with minimal manual correction. In tumor necrosis factor or TNF transgenic mice, the deep learning method maintained high performance despite progressive inflammatory erosive disease, with the accuracy of around 98%during the early disease progression stages.

Although the model was trained exclusively on hindpaws, it could be applied directly to forepaws without retraining to obtain anatomically coherent segmentations, demonstrating strong generalization to complex structures. It was found that the progression of severe erosive disease and joint destruction was accelerated in TNF transgenic female forepaws. Our protocol improves segmentation accuracy and automation, while still allowing manual corrections and continuous performance enhancements through added expert annotations.

Our future research aims to incorporate multiple unique structures and disease models and automated algorithms to continue enhancing flexibility and application.