Assessment of Anatomical Changes on Cone Beam Computed Tomography During Radiation Therapy Using TensorFlow-Based AI
Abstract
Purpose
This study developed an AI model to distinguish anatomical changes and quantify variations in patient alignment on cone beam computed tomography (CBCT) images between fractions of external beam radiation therapy treatments.
Methods
A TensorFlow based 3D-Unet Model was created and trained to recognize the differences between CBCT scans. Training data was generated by creating known augmentations that mimicked clinically relevant transformations. These transformations consisted of shifts, rotations, and scales of different magnitudes. Images were then paired by selecting two images from different fractions with known differences. The model was trained in three phases. Phase one trained the model to recognize segmentation and similarity, phase two focused on learning classification of transformations between the image sets, and phase three enabled estimation of transformation magnitude. Evaluation of the model was done using a specifically designed evaluation dataset.
Results
The model was trained on 1,409 pairs of CBCT scans for previously treated prostate patients. Training consisted of 10 epochs and an 80%/20% learning to validation split. Heat maps were generated and demonstrated near identical agreement for the true and predicted difference. The model had strong performance in segmentation with DICE coefficients consistently around 0.90 and as high as 0.97. The predicted similarity score showcased the model’s capability to detect and flag large anatomical changes. It was able to successfully identify transformations 85.8% of the time as well as quantify the magnitude of the transformation 73.4% of the time during training.
Conclusion
An AI model with these capabilities has the potential to automate intrafraction review of patient alignment and anatomy. The model provides a visual representation along with metrics that give Physicians, Medical Physicists, and Radiation Therapists, quantitative information to make clinical decisions. Further expansion and training of the model has potential to guide adaptive workflows during patient treatments.