Evolved size-specific dose estimates for patient-specific organ doses from chest CT scans based on hybrid patient size vectors.

Abstract

This study aims to develop a neural network-based method for predicting patient-specific organ doses from chest CT scans, utilizing hybrid patient size vectors for enhanced computational efficiency, accuracy, and generality. A dataset of 705 chest CT scans was retrospectively analyzed to construct predictive models for organ dose estimation. The proposed approach employs high dimensional hybrid vectors to represent patient size, combining muti-slice parameters regarding lateral dimension, anteroposterior dimension, and water-equivalent diameter (D_w). These vectors are used to train fully-connected neural networks, which are designed to correlate high-dimensional patient size features with reference organ doses obtained from Monte Carlo simulations. The performance of the neural networks was evaluated using separate test cohorts, with metrics such as mean absolute percentage error (MAPE) and coefficient of determination (R²) to evaluate predictive accuracy and generality. For the left lung, right lung, heart, and spinal cord, the trained neural networks respectively achieve MAPE values of 2.94%, 2.79%, 7.04%, and 6.76%, and R² values of 0.98, 0.99, 0.93, and 0.91. The maximal discrepancy between reference and predicted values is less than 10% for the left and right lungs, and less than 20% for the heart and spinal cord. With 5-fold cross-validation, the maximal perturbation does not exceed 1% in MAPE and 0.05 in R². By incorporating hybrid patient size vectors, the neural network models achieve superior accuracy in organ dose estimation compared with traditional size specific dose estimates, paving the way for online swift organ dose screening in clinical practice.