Swift prediction of personalized head and chest organ doses from CT examinations via neural networks with optimized quantity of hidden layers and radiomics features

Objective

To utilize radiomics features to enhance the prediction of personalized organ doses from CT scans, in order to explore methods for improving neural network-based models.

Methods

Patient CT DICOM files were processed using DeepViewer to define regions of interest (ROIs) in their organs. Radiomics features were extracted from the CT images and ROIs, and benchmark organ doses were calculated using Monte Carlo simulations. Fully-connected neural networks (FCNN) were trained with radiomics features to predict organ doses. The FCNN model was optimized by adjusting the number of input radiomics features and FCNN layers. Performance was evaluated using relative root mean squared error (RRMSE) and R-squared (R²).

Results

Higher RRMSE and lower R² values are observed when fewer than 30 input radiomics features are used for head CTs and fewer than 10 for chesst CTs. Increasing input features didn't significantly improve FCNN's performance. For head CTs, FCNN's layer quantities affected predictive stability, with better robustness observed with 4- and 5-layer FCNN. Specifically, the median RRMSE was reduced to 8.14% for the brain, 10.27% for the left eye, and 10.16% for the right eye when using 30 or more radiomics features. For chest CTs, the model's predictive stability was less sensitive to the number of layers, with median RRMSE values of 9.58% for the left lung and 9.44% for the right lung, and R² values of 0.76 for both lungs.

Conclusions

Optimizing feature quantities and neural network layers enhances performance in predicting organ doses from CT scans. Specifically, head CTs show optimal results with 4–5 layers, while chest CTs do not significantly benefit from increased layers.