Deep learning model mapped cervical spinal cord injuries in 711 patients

A research team developed a new AI architecture to help doctors evaluate spinal cord injuries more reliably. The system analyzes MRI scans to identify the exact location and extent of damage.

Traditionally, doctors assess these injuries by looking at scans, a process that can be subjective. Factors like small lesion sizes and blurry boundaries often lead to different interpretations among medical experts.

To train this tool, the researchers used a multi-center dataset containing 711 patient cases. The software uses a technique called Monte Carlo dropout to measure its own uncertainty, which helps clarify ambiguous lesion boundaries.

In testing, the model achieved a 70.44% accuracy rate for segmenting injuries. It also produced a 3D map that pinpoints the level of compression with high precision, mapping the lesion's length to the correct part of the spine in about 80.8% of cases.

Beyond just visualizing the injury, the software links imaging data to physical symptoms, such as the patient's motor control and impairment scores. This provides a clearer standard for tracking how a patient recovers over time.

While the model performs well at standardizing assessments, it is currently designed for research and monitoring purposes. The study does not show how it would perform in a live emergency room setting, nor does it replace the final clinical judgment required for complex surgical decisions.