Edward Wang: Hi there, my name is Edward, and I'm an MD/PhD candidate at Western University in Ontario, Canada, and today, I'd like to share with you a little bit about our recent work in the Red Journal, in which we developed an artificial intelligence model for real-time dosimetry prediction of complex multi-lesion lung SABR plans. As the number of lesions being treated with SABR grows, such as in the setting of oligometastatic disease, plans become more complex and it becomes increasingly challenging for radiation oncologists to choose a proper prescription or even to determine whether or not treatment is possible. The lungs are a particularly challenging site because lesions of varying sizes can be present anywhere within the lungs and there is a plethora of prescription options to choose from. The goal of this work was to develop a lung SABR prediction model that allows radiation oncologists to view in real-time the dosimetry of a multi- lesion lung SABR plan, prior to initiating the complex, lengthy and resource intensive treatment planning workflow. Now, I'd like to show you how radiation oncologists can use our model by walking through an example case. Here, we've integrated our model into an extension for 3-D Slicer, a free and open-source image computing platform. Our example case is a patient with 10 lesions from the COMET-10 trial. All lesions were treated with 35 Gy in 5 fractions, and the dose distribution that you see here is from the delivery plan. COMET-10 used relatively conservative prescriptions and now we'll use your model to determine if dose escalation was possible. First, we'll start with 40 Gy in 5 fractions. For the first case, I will let the model run without cutting the video, so you can see how long it takes from beginning to end. Our model uses the patient's planning CT scan, structure set, and a simple, initial estimate of dose to predict the 3-D dose distribution. We developed our model from 125 plans from patients who received multi-lesion lung SABR at our institution, with 80% of the plans used for training and 20% for evaluation. After our model was created, we extended our data collection period by over two years and collected an additional 90 multi-lesion plans for out of sample validation. Now that computation is finished, you can see the dose distribution present in the three viewing planes. Additionally, on the left-hand side, you can see how the DVH metrics of the prediction compares to the clinical constraints, which are adapted from the ones used at our institution and converted to EQD2. As you can see, after escalating the dose to 40 Gy in 5 fractions, all constraints are still met. Now, let's escalate the dose further, up to 55 Gy in 5 fractions. As you can see, even at this prescription level, all OAR constraints are met for this patient. For the purposes of demonstration, I've used the same prescription for every lesion, but you can also set a different dose and fractionation for each individual lesion. We hope that this work will be useful to the radiation oncology community. Please reach out if you have any questions and thank you very much for watching.
