Machine learning for optimizing mAs in KUB radiography with metal implants

Abstract

Background and purpose: Kidney-ureter-bladder (KUB) radiography is a common examination that exposes patients to a higher radiation dose and increased cancer risk; therefore, it is important to estimate suitable exposure factors for each patient prior to radiography. The present study aimed to utilize machine learning (ML) approach to predicting the suitable milliampere-seconds (mAs) and reducing overexposure in patients with metal implants during KUB radiography.

Methods: A phantom was used to understand the effect of metal implants on radiation exposure during KUB radiography with automatic exposure control (AEC) technique. Subsequently, we retrospectively enrolled 619 subjects, including 56 with metal implants and 563 without, from one hospital (group A) and 323 subjects, including 89 with metal implants and 234 without, from another hospital (group B). All subjects underwent both KUB radiography and physiological examinations on the same day. Data on body parameters and exposure factors were retrieved from hospital database. To train the prediction model, the dataset of group A without metal implants was randomly divided into 80% and 20% for training and testing sets, respectively. Five different ML algorithms were utilized to train the prediction model using 10-fold cross-validation. The correlation coefficients (CC), mean average error (MAE), normalized root mean squared errors (nRMSE), and R-square (R²) were compared to find the optimal model. For external validation, the dataset of group B was randomly separated into 80% and 20% for training and testing sets, respectively. The training sets of both groups were combined for transfer learning, and the testing set of the group B was used to assess the optimal model. Furthermore, the final model was utilized to predict an appropriate mAs for patients with metal implants in both groups. Statistical analysis was performed to understand the differences between datasets, phantom settings, and ML models. Comparisons were considered significance if p < 0.05.

Results: The phantom experiment demonstrated that the metal plate significantly increased the mAs and reached exposure (REX) values when using AEC technique during KUB radiography. The comparison of patient data showed that the patients with metal implants had significantly higher mAs and REX than those without in both groups. In group A, the ML comparisons showed that the artificial neural network (ANN) model outperformed other ML models in predicting mAs based on the testing set, exhibiting the highest CC of 0.791 ± 0.007 and R² of 0.6193 ± 0.010. In group B, the external validation based on transfer learning demonstrated that the ANN model achieved the CC of 0.837 ± 0.051 and R² of 0.823 ± 0.007 in the testing set. For patients with metal implants, the ANN model-predicted mAs was significantly lower than those obtained using AEC technique in both groups.

Conclusion: We concluded that the ML approach is suitable for building the model for predicting appropriate mAs and reducing overexposure in patients with metal implants during KUB radiography.

Keywords: KUB radiography; mAs; machine learning; target exposure indicator.