Optimizing U-Net Architecture with Feed-Forward Neural Networks for Precise Cobb Angle Prediction in Scoliosis Diagnosis

Abstract

In the burgeoning field of Artificial Intelligence (AI) and its notable subsets, such as Deep Learning (DL), there is evidence of its transformative impact in assisting clinicians, particularly in diagnosing scoliosis. AI is unrivaled for its speed and precision in analyzing medical images, including X-rays and computed tomography (CT) scans. However, the path does not lack obstacles. Biases, unanticipated outcomes, and false positive and negative predictions present significant challenges. Our research employed three complex experimental sets, each focusing on adapting the U-Net architecture. Through a nuanced combination of feed-forward neural network (FFNN) configurations and hyperparameters, we endeavored to determine the most effective nonlinear regression model configuration for predicting the Cobb angle. This was done with the dual purpose of reducing AI training time without sacrificing predictive accuracy. Utilizing the capabilities of the PyTorch framework, we meticulously crafted and refined the deep learning models for each of the three experiments, focusing on an FFFN dropout rate of p=0.45. The Root Mean Square Error (RMSE), the number of epochs, and the number of nodes spanning three hidden layers in each FFFN were utilized as crucial performance metrics while a base learning rate of 0.001 was maintained. Notably, during the optimization phase, one of the experiments incorporated a learning rate scheduler to protect against potential pitfalls such as local minima and saddle points. A judiciously incorporated Early Stopping technique, triggered between the patience range of 5-10 epochs, ensured model stability as the Mean Squared Error (MSE) plateau loss approached approximately 1. Consequently, the model converged between 50 and 82 epochs. We hypothesize that our proposed architecture holds promise for future refinements, conditioned on assiduous experimentation with an array of medical deep learning paradigms.