1. How can deep learning approaches handle transitional vertebrae in vertebrae segmentation and identification?
In this work, we investigate the combination of prior knowledge on the full structure of the spine with a deep learning approach to handle transitional vertebrae in vertebrae segmentation and identification. We propose to iteratively cycle between the localization, segmentation, and identification tasks, which uses deep networks, while enforcing anatomical consistency with statistical priors. The priors are used to localize vertebrae by leveraging learned statistics of vertebrae volume and inter-vertebral distances, which exhibit more robustness to pathological cases than shape or appearance models. For the identification task, we encode the admissible configurations in a graphical model that leverages local deep-network predictions and aggregates them into an anatomically consistent result. Our experiments demonstrate that this strategy successfully handles local inaccurate predictions and performs better than other methods with transitional vertebrae while providing state-of-the-art results on standard benchmarks such as VerSe20. We also propose to learn statistics conditioned on the spine level, compute adaptative thresholds that automatically vary across spine levels and patients, and use a two-step approach to localize and identify vertebrae. This approach explicitly models transitional vertebrae and improves the performance of deep learning methods in handling them.
read more
2. How does the Anatomic Consistency Cycle enforce anatomic consistencies?
The Anatomic Consistency Cycle enforces anatomic consistencies by cycling through localization, segmentation, and identification tasks. It starts by segmenting a spine mask, which enables locating individual vertebrae. Vertebrae segmentation masks are estimated and refined iteratively. Identifications are made using locations and segmentation masks. The obtained identifications are used to enforce finer anatomic consistency constraints and detect new candidate locations. The process continues until the proposed consistency criteria are satisfied or the set of detected locations, segmentation masks, and identifications do not change. Any remaining inconsistencies are reported in the results. This cycle ensures anatomic consistencies among the tasks, leading to accurate vertebrae identification and segmentation.
read more
3. What are anatomical constraints in spine segmentation?
Anatomical constraints in spine segmentation involve leveraging deep networks for localization and segmentation, using residual connected components, vertebrae volume statistics, and inter-vertebral distance models. These constraints help identify individual vertebrae and ensure accurate segmentation. The process starts with segmenting the spine from a 3D CT scan, followed by identifying the location, segmentation, and identification of each vertebra. The residual connected component, which represents the whole spine mask at the first cycle, is considered. Vertebrae volume statistics are used to determine if a residual is a vertebra or noise, with a threshold of 50% of the predicted size. The inter-vertebral distance constraint utilizes statistical models, including Gaussian distributions and linear regressors, to detect abnormally large distances between vertebrae. Additionally, identification is used to check if the spine extremes are complete, with new candidate locations added if necessary. Overall, anatomical constraints play a crucial role in achieving accurate spine segmentation.
read more
4. How does the global reasoning approach improve vertebrae identification?
The global reasoning approach improves vertebrae identification by enforcing the natural ordering of vertebrae, ensuring consecutive vertebrae have consecutive labels. It uses a graph optimization strategy, where a graph with n x 24 nodes is created, representing n consecutive vertebrae and 24 vertebra classes. Unary costs are populated with individual local probabilities, and binary costs are added to prevent group swapping. Edges between nodes enforce consistent classes for consecutive vertebrae. Special attention is given to transitional vertebrae, with additional edges and higher costs. The shortest path search algorithm is used to find the optimal set of labels, and post-processing adjusts repeated instances of T12 or L5 to T13 or L6, respectively. This global approach enhances the accuracy of vertebrae identification by considering the context and ordering of vertebrae in the spine.
read more