22nd AIAI 2026, 16 - 19 July 2026, Chania, Crete, Greece

Automated Scoliosis Assessment via Dual-Stage YOLOv8 Detection and Attention U-Net Segmentation

Hayat Dvir, Abbassov Refael, Weiss Cohen Miri

Abstract:

  This work presents an automated dual-stage deep learning framework for vertebral segmentation and scoliosis assessment from CT images, combining YOLOv8 object detection with Attention U-Net segmentation. YOLOv8 first localizes individual vertebrae by generating precise bounding boxes that isolate regions of interest, which are then processed by the Attention U-Net, whose attention gates suppress irrelevant background while enhancing salient vertebral features for accurate boundary delineation. Trained on a manually annotated CT dataset, the framework achieves near-perfect detection precision and recall, a Dice coefficient of 0.97, and an IoU of 0.94. The system further extracts vertebral centroids to construct a linear spinal representation, enabling quantitative curvature analysis and automated scoliosis severity classification, offering a robust and fully automated tool for clinical monitoring and treatment planning.  

*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.