Room: Exhibit Hall | Forum 3
Purpose: Due to the use of kV photon beams in small animal radiotherapy, the dose deposition is strongly dependent on the elemental tissue composition. We present a novel method to extract continuous elemental tissue compositions from dual energy CBCT (DE-CBCT) images.
Methods: We use principal component analysis (PCA) to reduce the dimensionality of the material space. For each DE-CBCT image, we first classify individual voxels as air, soft tissue, or bone through a support vector machine. In a next step, a shallow neural network determines the weights of the principal components. All images were acquired on the Small Animal Radiation Research Platform (XStrahl Ltd). The DE-CBCT imaging protocol was optimised over a range of energy and tin filter combinations. The model is calibrated on a mouse-sized solid water phantom with 10 tissue-equivalent material inserts with known elemental composition. We validate the model on a second phantom with 6 different tissue-equivalent inserts, and compare the predictions to a conventional PCA method with a linear regression model. We additionally apply the new method to a mouse DE-CBCT.
Results: The material composition in the 6 inserts of the validation phantom was extracted with a mean squared error of 1%, averaged over all inserts. This is an improvement over conventional PCA methods without a neural network (3% average mean squared error), especially for bony tissues. While no ground truth is available for the mouse tissues, we found good agreement with the composition from the material database, for instance within 3% for the mouse brain.
Conclusion: We presented a novel method to extract the elemental tissue composition from DE-CBCT images. The continuous assignment of materials is expected to reduce discretisation errors in dose calculations compared to current approaches, where only 4 to 7 materials are used.