Online target volume estimation and prediction from an interlaced slice acquisition - A manifold embedding and learning approach

In radiotherapy it is critical to have access to real-time volumetric information to support online dose calculation and motion management. MRI-guidance offers an online imaging platform but is restricted by image acquisition speed. This work alleviates this limitation by integrating processing techniques with an interlaced 2D real-time acquisition protocol. We characterize the volumetric anatomical states as samples on a manifold, and consider the alternating 2D slice acquisition as observation models. We infer sample locations in the manifold from partial observations and extrapolate on the manifold to generate real-time target predictions. A series of 10 adjacent images were repeatedly acquired at three frames per second in an interleaved fashion using a 0.35 T MRI-guided radiotherapy system. Eight volunteer studies were performed during free breathing utilizing normal anatomical features as targets. Locally linear embedding (LLE) was combined with manifold alignment to establish correspondence across slice positions. Multislice target contours were generated using a LLE-based motion model for each real-time image. Motion predictions were performed using a weighted k-nearest neighbor based inference with respect to the underlying volume manifold. In the absence of a 3D ground-truth, we evaluate the part of the volume where the acquisition is available retrospectively. The dice similarity coefficient and centroid distance were on average 0.84 and 1.75 mm respectively. This work reports a novel approach and demonstrates promise to achieve volumetric quantifications from partial image information online.

Duke Scholars

Author John Stephen Ginn Radiation Oncology