Room: Exhibit Hall | Forum 8
Purpose: The delineation of gross tumor volume (GTV) for glioma patients is a laborious process that involves visual inspection of two MR image sequences, T1 and T2 FLAIR, acquired both pre- and post-operatively, and a planning CT sequence. Our goal is to develop a systematic approach for GTV delineation using a deep learning method that combines all the input image sequences in 3D with a few manually delineated 2D slices. Our hypothesis is that the information found in these five image sequences together with a small portion of user supplied 2D information will generate high-quality 3D delineations.
Methods: Pre- and post-operative MR-T1 and MR-T2 FLAIR images, and a planning CT image were collected for 200 glioma patients previously curatively treated at our institution with intensity modulated radiation therapy (IMRT). For each case, the GTV was delineated by a treating physician on CT/MR fusion using pre- and post-operative T1 and T2 FLAIR sequences. We developed a 3D convolutional neural network based on a U-net architecture for automated delineation of the GTV. The model takes all of the image sequences in 3D as input together with the user supplied 2D delineations. The output of the model is a full 3D delineation.
Results: The dataset was split into training (80%) and validation (20%) sets and the model was trained on a high-performance computer. Model performance improved as measured by the Dice-Sorensen coefficient (DSC) from 0.65 to 0.90 when three of the ground truth slices were provided as input.
Conclusion: The interactive deep learning method proposed in this study can be implemented as a tool to assist radiation oncologist to delineate post-operative glioma regions. It has the potential to reduce inter-observer variability and free up time for delineating physicians.
Funding Support, Disclosures, and Conflict of Interest: Part of the research is funded by RaySearch Laboratories.