Neural network-based multi-task learning to assist planning of posterior spinal fusion surgery for adolescent idiopathic scoliosis.

Abstract

Purpose: Posterior spinal instrumentation and fusion (PSF) is the gold standard for severe adolescent idiopathic scoliosis (AIS), yet instrumentation strategies vary widely, often leading to suboptimal results. Deep learning's potential in AIS planning is underexplored.

Methods: This study trained and validated an artificial neural network multi-task learning model (NNML) using preoperative clinical and radiographic data from 189 AIS patients with Lenke 1A and 2A curves enrolled in the MIMO Clinical Trial (NCT01792609). The model mimics experienced spine surgeons' decision-making for selecting the upper and the lower instrumented vertebrae (UIV, LIV), determining rod curvature, and predicting screw density based on the study's randomized allocation. Models were trained with data from 179 patients, utilizing tenfold cross-validation, and externally validated on 10 patients from a separate hospital and surgeons outside the training set. For UIV and LIV selection, accuracy within the top two predictions was used as a classification performance metric, ensuring that other clinically relevant alternatives were considered.

Results: The NNML, which comprised 83 inputs and multiple hidden layers, led to significant gains over ST-NN and proved more robust during the internal validation (loss 6.2 vs. 9.3; p ≤ 0.01). It showed 82-95% and 80-100% accuracy for UIV and LIV predictions and 70-90% accuracy for predicting the rod curvatures ± 5°. The RMSE for the screw density and rod curvature predictions was 0.2-0.3 and 3.7-5.6°, respectively.

Conclusion: An NNML can better use the features of relevant AIS patients for mixed task prediction pertinent to PSF surgery planning than ST-NN. In addition, NNML was capable of mimicking experienced spine surgeons' decision-making process when designing the instrumentation.