Optimizing deep learning-based vertebra identification on CT images for efficient clinical deployment

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] en sciences informatiques, à finalité spécialisée

Abstract

enIdentifying and numbering vertebrae on CT images is a crucial preliminary step in surgical planning and in the monitoring of spinal pathologies. However, for radiologists, this task is repetitive, time-consuming and prone to errors, particularly given the increasing workload and time pressure in clinical settings. Although numerous deep learning methods have been proposed to automate vertebra identification, the vast majority are designed in a research context and rely on architectures and pipelines that are too slow for real-world clinical deployment. This master thesis develops a deep learning model capable of automatically localizing each vertebra on CT images and assigning it its correct anatomical label, while satisfying the performance and efficiency constraints required for a future integration into a commercial PACS solution. To achieve this, we build on nnLandmark, a recent fork of the well-established nnU-Net framework, specialized in landmark detection through heatmap regression. We produced a thorough time breakdown of the nnLandmark pipeline, which revealed that the actual neural network inference represented only a small fraction of the total execution time, with preprocessing and postprocessing being the true bottlenecks. Based on this analysis, several optimizations were introduced: single-threaded execution for single-image inference, GPU-accelerated resampling, direct centroid extraction in the network space through coordinate scaling rather than resampling the full prediction volume, and model conversion to ONNX and TensorRT formats. Combined with batch size tuning and a center-of-mass-based centroid extraction, the model achieves an identification rate of 93.62\%. On the efficiency side, the optimized pipeline achieves an average processing time of 1.40 seconds per image on a server-grade GPU and 1.58 seconds on a consumer-grade GPU, well within the few-second requirement for clinical deployment.