Identification of Atypical Scoliosis Patterns Using X-ray Images Based on Fine-Grained Techniques in Deep Learning

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Spine Surgery

Background:

• Scoliosis is a complex spinal deformity.
• Atypical scoliosis patterns can be associated with underlying neurological conditions like Chiari Malformation type I (CMS).
• Accurate and early screening for these conditions is crucial for effective management.

Purpose of the Study:

• To develop a deep learning-based classification model using X-ray images.
• To screen for scoliosis and specifically identify atypical patterns associated with CMS.
• To evaluate the performance of different ResNet-50 model configurations for this diagnostic task.

Main Methods:

• Retrospective analysis of 508 pairs of coronal and sagittal X-ray images from patients with CMS, adolescent idiopathic scoliosis (AIS), and normal controls (NC).
• Development and evaluation of multiple ResNet-50 deep learning models (Coronal, Sagittal, Dual, Concat, Bilinear).
• Performance assessment using accuracy, sensitivity, specificity, PPV, NPV, ROC curves, and heatmaps for both scoliosis and CMS diagnosis.

Main Results:

• The ResNet-50 Coronal model showed the best performance for general scoliosis screening.
• For CMS diagnosis, ResNet-50 Coronal and ResNet-50 Dual models performed optimally.
• The ResNet-50 Dual model achieved performance comparable to senior spine surgeons, while the ResNet-50 Coronal model surpassed them in specificity and PPV for CMS detection. Heatmaps highlighted key features like atypical curves and trunk tilt.

Conclusions:

• A fine-grained classification model based on ResNet-50 can accurately screen for atypical scoliosis patterns associated with CMS.
• The model effectively utilizes radiographic features such as curve type, lateral shift, segment length, and trunk tilt for classification.
• This deep learning approach offers a promising tool for early detection and diagnosis of CMS-related spinal deformities.