Comprehensive Summary
This article aims to evaluate the limitations of radiographic imaging and single-task learning (STL) models in adolescent idiopathic scoliosis assessment by use of a combination of multi-task learning (MTL) models and optical imaging for enhanced accuracy. To collect data, 160 individuals aged 7-18 years with radiographically confirmed AIS were selected, excluding patients with neuromuscular, syndromic, or congenital scoliosis. A multi-task deep learning model was trained to predict Cobb angle, curve type (thoracic, lumbar, mixed, none), and curve direction (left, right, none) from shared morphological features. The MTL model outperformed the STL model in detecting Cobb angle with a lower average mean absolute error (MAE) of 2.9 degrees versus the STL model’s 5.4 degrees. The MTL also achieved higher accuracy across three of the four classes of curve types against the STL model: lumbar (89% vs 65%), mixed (89% vs 75%), none (75% vs 66%). The one area MTL performed worse than STL was thoracic (73% vs 82%). For curve direction accuracy, STL performed decently, lacking on right curve direction detection: 87% for left, 97% for none, and 66% for right. MTL was more balanced, with 75% for left, 80% for none, and 80% for right. Overall, the MTL model performed well and, in most cases, better than the STL model.
Outcomes and Implications
Traditional methods of spinal imaging and diagnosis can be harmful to patients, time-consuming, or inaccurate. The STL model was introduced as a possible harmless alternative, however this model is highly limited. Multi-task deep learning systems provide a safe, fast, and radiation-free alternative to an X-Ray, allowing for much faster and more accurate diagnoses. Future models can be trained with larger datasets for independent validation, which would increase accuracy and increase the clinical applicability of this model.