(Wednesday, 7/17/2019) 7:30 AM - 9:30 AM
Room: 225BCD
Purpose: To develop a deep learning algorithm that predicts volumetric changes of esophagus during radiotherapy of locally advanced lung cancer and signals an adaptive replanning in the scenario of volume expansion, therefore reducing esophagus dose and the possibility of acute esophagitis which is one of the major toxicities associated with lung radiotherapy.
Methods: The deep learning algorithm, EP-net, utilized the first three weekly T2-W MRI scans to predict volumetric changes of esophagus in the following three weeks of radiotherapy (total 6 weeks). We adopted a patch (55x55mm) based design to accommodate hundreds of thousands of esophagus changing patterns observed on an IRB-approved clinical protocol including 9 patients. EP-net started with extracting features of an esophagus patch using a convolutional neural network. Subsequently an integrated recurrence neural network analyzed the extracted features and generated a prediction of spatial esophagus distribution on the patch level. Finally, a reconstruction model unified the patched prediction and assembled an overall look of esophagus. Comparing the predictions to experts-contoured esophagus, we calculated Dice, root mean square surface distance (RMSSD) and the accuracy of predicted volume changes to evaluate the performance of EP-net under a leave-one-out scheme.
Results: The volumetric changes of esophagus ranged from -7% to 59% during the last three weeks of lung radiotherapy. EP-net predicted the spatial distribution of esophagus for the last three weeks with a Dice of (0.84±0.04, 0.83±0.04, 0.81±0.06), RMSSD of (1.2±0.2mm, 1.3±0.1mm, 1.4±0.3mm), respectively. The correlation between the predicted and actual esophagus volume was 0.98, while the accuracy of predicting volume expansion was 0.78.
Conclusion: EP-net can predict the volumetric changes of esophagus in a longitudinal imaging study of lung radiotherapy. Its design is fully compatible with the clinical workflow, and its timely prediction is set to play a critical role in the decision-making of adaptive radiotherapy.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by Varian Medical Systems.
Keywords
MRI, Image Processing, Radiation Therapy
Taxonomy
IM- MRI : Quantitative imaging/analysis
Contact Email
