Deep learning can be used to predict patient eye lens radiation dose during neuro-interventional radiology procedures, according to a study presented Saturday at RSNA 2020.
Such a technique could help prevent the development of radiation-induced cataracts in patients under these procedures.
While skin dose has been the primary concern in neuro-interventional procedures because of the potential for radiation-induced skin injuries, these procedures also have the potential for a high dose to the patient's eye lens, explained Jacob Collins, MS, a PhD student at the University at Buffalo, State University of New York. In fact, the dose often exceeds 500 mGy, the amount estimated by the International Commission on Radiological Protection to induce cataracts.
Collins and his colleagues had the idea to calculate patient lens dose using deep learning methods.
"At its base, the problem is to predict a dependent variable — lens dose — given a set of independent variables — geometric parameters," Collins said. "Due to the type and problem complexity, a dense neural network was chosen to provide accurate answers."
They used Monte Carlo simulations to create their ground truth data, using a number of different data points: entrance field size, gantry angulation, head shift, and left eye or right eye.
The researchers came up with a dataset of 1,545 data points, split into a training set of 1,283 samples, a validation set of 143, and a testing set of 119.
Collins and his colleagues used K-fold cross validation, where the data was split into folds of 10. One fold is used as a validation set, while the remaining nine are used for training. This continues until each is used as a validation set, resulting in 10 different models with unique validation sets.
"Our best accuracy was achieved by taking a combination of models to get a final prediction," Collins said. "And with that method we got a mean absolute percentage error (MAPE) of 7.8%."
A MAPE of 0 would be perfect agreement, and the team's goal was to stay below 10%. The team agreed less than 8% was a good sign they were on the right track.
"I've been perfecting the model and introducing more parameters," Collins said. "Once the deep learning model is found to be accurate and work well, the next step is to actually integrate the model into the dose tracking system."
Next, the tracking system records the geometric parameter and the exposure parameters while the procedure is taking place.
"We can feed those into the deep learning network. It can predict the eye lens dose and be displayed for the staff to see," Collins said. "And it would let them know whether they are getting close to the 500 mGy threshold so they could possibly move the patient into a different position, still achieve the clinical task, and save the cataracts."
For More Information:
View the RSNA 2020 session, Investigation of Using Deep Learning to Predict Patient Eye-Lens Dose During Neuro-interventional Procedures — SSIN05 at RSNA2020.RSNA.org.
