Estimation of the radiation dose in pregnancy: an automated patient-specific model using convolutional neural networks

Objectives The conceptus dose during diagnostic imaging procedures for pregnant patients raises health concerns owing to the high radiosensitivity of the developing embryo/fetus. The aim of this work is to develop a methodology for automated construction of patient-specific computational phantoms based on actual patient CT images to enable accurate estimation of conceptus dose. Methods We developed a 3D deep convolutional network algorithm for automated segmentation of CT images to build realistic computational phantoms. The neural network architecture consists of analysis and synthesis paths with four resolution levels each, trained on manually labeled CT scans of six identified anatomical structures. Thirty-two CT exams were augmented to 128 datasets and randomly split into 80%/20% for training/testing. The absorbed doses for six segmented organs/tissues from abdominal CT scans were estimated using Monte Carlo calculations. The resulting radiation doses were then compared between the computational models generated using automated segmentation and manual segmentation, serving as reference. Results The Dice similarity coefficient for identified internal organs between manual segmentation and automated segmentation results varies from 0.92 to 0.98 while the mean Hausdorff distance for the uterus is 16.1 mm. The mean absorbed dose for the uterus is 2.9 mGy whereas the mean organ dose differences between manual and automated segmentation techniques are 0.07%, -0.45%, -1.55%, -0.48%, -0.12%, and 0.28% for the kidney, liver, lung, skeleton, uterus, and total body, respectively. Conclusion The proposed methodology allows automated construction of realistic computational models that can be exploited to estimate patient-specific organ radiation doses from radiological imaging procedures.