Poster Presentation 50th International Society for the Study of the Lumbar Spine Annual Meeting 2024

SpineGPM: A generative pretrained model for 3D spine synthesis based on back geometry with single-centre clinical validation on scoliosis diagnosis (#245)

Teng Zhang 1 , Nan Meng 1 , Moxin Zhao 1 , Wenting Zhong 1 , Yong Hai 2 , Jason Pui-Yin Cheung 1 , Ashish Diwan 3
  1. Digital Health Laboratory, Li Ka Shing Faculty of Medicine, The University of Hong Kong, Pokfulam, Hong Kong, China
  2. Department of Orthopaedics and Traumatology, Capital Medical University Affiliated Chaoyang Hospital, Beijing
  3. Spine Labs, St George and Sutherland Clinical School, University of New South Wales, Sydney, NSW, Australia

 

Introduction

Adolescent Idiopathic Scoliosis (AIS) is a common three-dimensional spinal deformity in children. If not addressed promptly, the condition may progress rapidly. Traditional treatment methods require spinal X-ray examinations and regular follow-ups which will increase the radiation exposure. Therefore, the assessment of spinal morphology using three-dimensional optical techniques is crucial for more frequent monitoring of disease progression. We have developed a novel 3D spine generation model (SpineGPM) that utilizes patient surface 3D geometry to first generates and reconstructs the 3D spine model. This approach may potentially circumvent the risks associated with repetitive radiographs, while easily accessible and cost efficient. This study aims to prospectively validate SpineGPM generated 3D spine model with an independent centre.

 

Method

A total number of 3983 AIS patients who attended two local clinics (Queen Mary Hospital and Duchess of Kent Children’s Hospital at Sandy Bay) were recruited to form the dataset used for model development and evaluation. A prospective collection of 302 patients was used for testing. For each patient, 3D images of the patient’s unclothed back to generate the 3D spine model. A spine specialist quantifies the deformity severities using the SpineGPM generated 3D spine model, while blinded to the ground truth obtained by the whole spine radiograph collected on the same day.  Sensitivity, specificity and negative prediction value (NPV) of the deformity severity based on the SpineGPM generated 3D spine model were assessed.

 

Results

In the prospective cohort, there were 85 individuals (28.1%) identified as normal-mild, 184 individuals (60.9%) identified as moderate and the rest 33 individuals (10.9%) identified as severe. For severity assessments, the SpineGPM generated 3D spine model achieved sensitivity=0.835, specificity=0.954, NPV=0.937 among 85 normal-mild patients, and sensitivity=0.935, specificity=0.856, NPV=0.894 among 184 moderate patients, and sensitivity=0.909, specificity=0.993, NPV=0.989 among 33 severe patients.

 

Discussion

The single centre prospective testing revealed the SpineGPM can achieve comparable results with the radiographs when assessing the spine deformity in children. However, the prospective data volume in this experiment is limited, and the results obtained require further in-depth multi-centre validation. In subsequent research, we will actively collaborate with national and international centres to conduct multi-centre validation.

 

655582a23fca7-Weixin+Screenshot_20231116105351.png

Figure 1. (a) confusion matrix comparing the severity classification based on the SpineGPM generated 3D spine model and the radiograph. (b) visualization of the SpineGPM generated 3D spine models for mild, moderate and severe deformities (from left to right).