Room: Exhibit Hall | Forum 2
Purpose: The expanding adoption of magnetic resonance imaging (MRI)-guided radiation therapy (MR-IGRT) has highlighted several clinical needs related to applications of MR-IGRT. Among these is automated segmentation of tissues and structures for use in a rapid, MR-only RT workflow. Eliminating CT simulation in favor of an MR-only workflow in which MRI is the sole imaging modality used for treatment planning and setup avoids several issues related most notably to multi-modality registration inaccuracies. Such a workflow lends itself well to a streamlined treatment process that is further enabled by the automated segmentation of structures relevant to treatment planning. We present here the application of our deep spatial pyramid convolutional framework to the semantic image segmentation problem for MR-only RT of the breast.
Methods: Segmentation of the left and right breasts, lungs, and heart is achieved using a novel convolutional framework we have developed. The framework utilizes atrous convolution applied at increasing rates in parallel at the interior of a generative model to effectively encode contextual information from multiple fields of view. A discriminative model is used to train the generator in an adversarial fashion. 1.5 T MR images acquired using an mDixon sequence acquired during MRI simulation are used as the basis for our automated segmentation.
Results: Test results demonstrate good agreement between the automatically generated contours and those drawn in the clinic. This is indicated quantitatively using the dice coefficient, with scores of 0.96, 0.95, 0.85, 0.83 and 0.86 for the left and right lung, left and right breast, and heart, respectively.
Conclusion: Automated semantic segmentation with our deep spatial pyramid convolutional framework enables a rapid, streamlined MR-only RT workflow. Segmentation in this manner may be used to perform bulk density overrides or paired with methods of synthetic CT generation for true MR-only RT of the breast.