Beyond Accuracy: Multi-Level Ordinal Assessment of Lumbar Spine Degeneration from Multiplanar MRI Using the RSNA 2024 (LumbarDISC) Dataset and Condition-Specific ViTs

Lumbar spine degeneration assessment from magnetic resonance imaging (MRI) involves multiple anatomical conditions, spinal levels, and ordered severity categories, making conventional accuracy alone insufficient for a complete evaluation. In this study, we present a Beyond-Accuracy framework for automated lumbar spine degeneration assessment using the RSNA 2024 Lumbar Spine Degenerative Classification dataset. Five independent Vision Transformer (ViT) pipelines were developed for spinal canal stenosis (SCS), left and right neural foraminal narrowing (LFN/RFN), and left and right subarticular stenosis (LSS/RSS). Localized MRI crops centered on annotated spinal-level coordinates were used to classify severity as Normal/Mild, Moderate, or Severe. For each condition, 24 ViT architecture and hyperparameter configurations were explored across ViT-B/16, ViT-L/16, ViT-B/32, and ViT-L/32 backbones, followed by repeated-run stability analysis. The selected condition-specific models achieved approximately 82–92% validation accuracy, with SCS reaching the highest performance at approximately 92%. Conventional measures, including accuracy, precision, recall, and F1-score, were complemented by Ordinal Accuracy (OA), Condition-Level Ordinal Accuracy (CLOA), and three Patient-Level Ordinal Accuracy (PLOA) formulations to characterize severity ordering at global, condition, and patient levels. The results show that exact classification accuracy alone does not fully capture model behavior in structured severity prediction. Combining condition-specific ViT pipelines with multi-level ordinal-aware evaluation provides a more comprehensive framework for lumbar degeneration assessment and supports the development of more interpretable MRI-based decision-support systems.