Room: Karl Dean Ballroom B1
Purpose: To develop and evaluate a deep learning-based tumor auto-contouring algorithm for real-time tumor tracking using linac-MR.
Methods: We modified and implemented a specific kind of neural network known as U-net in our auto-contouring algorithm, which was responsible for tumor segmentation from its surrounding anatomy. U-net was chosen due to its ideal network structure capable of combining low resolution context information and high resolution details of the input image in pixel-wise classification. Six non-small cell lung cancer patients were imaged with 3T MRI at ~4 frames per second (2D sagittal plane, free breathing). For each patient, an expert delineated gold standard contours (ROI_std) of the lung tumor in 130 consecutive images. ROI_std were delineated on 3T images. However, to evaluate our algorithm in the 0.5T linac-MR environment, the 3T-acquired images were noise-degraded to reflect the worst-case image quality of lung tumor at 0.5T. For each patient, the first 30 ROI_std were used for algorithm training, and the remaining 100 ROI_std were used for auto-contouring validation. Once trained, our algorithm was applied to segment tumors from the remaining 100 pseudo-0.5T images, generating 100 ROI_auto per patient. In each image, the dice similarity index (DSI), Hausdorff distance (HD), and centroid position difference (Î”d_centroid) were calculated between ROI_std and ROI_auto to measure their similarity.
Results: We achieved 87 â€“ 92% DSI, 2.4 â€“ 3.9mm HD, and 0.9 â€“ 1.5mm Î”d_centroid from six patients' data. Our algorithm successfully segmented the shape of moving tumors that exhibit significant shape changes during breathing (e.g., partial omission of tumor), and/or are surrounded by challenging anatomy (e.g., diffuse boundary between tumor and adjacent tissue).
Conclusion: We developed a deep learning-based tumor auto-contouring algorithm, and evaluated its performance with in-vivo MR images. Our algorithm showed promising results (>87% DSI) suggesting the feasibility of auto-contouring moving tumors on MR images.