Purpose: Summary statistics and fitting/regression equations are typically provided in pediatric CT cohort studies for patient groups without or with poorly derived patient-specific estimates. An accurate organ dose estimation methodology may not be available for clinicians and usually requires resources and specific knowledge to use. Thus, this study investigates the feasibility and accuracy in using machine learning methods to create a generalized model to predict patient-specific organ dose with as few clinical parameters/features as possible for easily accessible dose estimation.
Methods: Previously estimated patient-specific organ and effective dose data (1252 anonymized patients) and corresponding clinical and scan parameters/features were processed with Python. Clinical parameters include protocol name, patient age, patient weight, patient lateral width. Scan parameters include total collimation, kVp, mean mAs, CTDIvol, pitch, scan start and end locations. Stratified random sampling in each protocol group with probability of 0.8 (reproducible random sampling state 1~10) separates training from test data. Non-parametric and parametric regression methods, K-Nearest-Neighbor (KNN) regression and Neural Network (NN) regression from Scikit-learn library, were used to separately create machine learning models to predict a vector of 33 organ and effective doses. Preliminary KNN model was configured with a typical k-factor of 5 and preliminary NN models with four layers of neurons and ReLU activation.
Results: Organ and effective doses for test dataset were successfully predicted using two combination of parameter sets and KNN/NN methods with mean coefficient of determinant no less than 0.78, and mean dose difference less than 22%.
Conclusion: Prediction of organ and effective dose for CT exams based on few clinical features using machine learning methods is feasible and reasonably accurate. Such models may offer more expedient access of dose information for clinicians and their patients. Optimization of machine learning parameters (e.g., k-factor, NN layers) and exploring prediction with fewer parameters will be investigated.