Tailored Loss Functions to Improve Vertebra Centroid Localisation and Classification in Sagittal Spinal Radiographs

Le résumé fourni par la source

A healthy spine is essential for a high-quality life, yet spinal musculoskeletal issues cause suffering and significant socioeconomic burdens. Osteoporotic vertebral fractures (OVFs), for example, increase risk of future fractures, morbidity, and mortality. While spine radiography is the gold standard for OVF detection, manual interpretation is time-consuming and challenging. Hence, this work proposes an automated tool for simultaneous vertebral centroid localisation and classification in sagittal spinal radiographs, streamlining the initial input required for semi-automated diagnostic tools and enhancing the efficiency of spine assessment. To aid the model’s learning despite the partial annotations and varying fields of view (FOV) in the datasets used, various loss functions were developed and tested to enhance performance and allow for further training on more complete annotations in the future. A total of four different loss functions, as well as a baseline model trained with standard MSE loss, were compared. The best-performing models improved vertebral ID accuracy by over 5% above the baseline model in both datasets. Future work will leverage the outputs of these models for vertebra segmentation and fracture detection in a fully automated pipeline.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.