Liver Tumor Motion Estimation Based on Respiration-Induced Skin Change Using Principal Component Analysis and Nearest Neighbor Method

Abstract

This paper proposes a liver tumor motion tracking approach based on respiration-induced skin change by using principal component analysis (PCA) and nearest neighbor estimation (NNE). The proposed method is abbreviated as PCA-NNE and is developed on the NNE in our previous work. The PCA-NNE and NNE are both estimation models by correlating internal target motion and its external surrogate. Their difference is that the surrogate for NNE is the marker attached on skin, while for PCA-NNE, the surrogate is the entire skin change. In PCA-NNE, PCA derives several principal components (PCs) from skin, and then the PCs were inputted into NNE to estimate target motion. Through this way, PCA-NNE simplify the selection of surrogate. Similar to NNE, PCA-NNE needs data to construct patient-specific model before application. To evaluate PCA-NNE, we use a digital torso phantom to create abdominal computed tomography (CT) images of a virtual patient. By setting different breathing parameters and locations of a lesion, we simulated 21 sets of respiration-correlated CT with a length of nearly 1 minutes, and derived synchronized liver tumor motion and skin change from them. For each set, the data during the first 2 ~ 3 respiration periods were used for modeling. The remaining data were used for evaluation. Two metrics were adopted for assessment. The root-mean-square error (RMSE) was used to validate its accuracy. The estimation time per sample (t) was used to evaluate its real-time performance. The results showed a RMSE of < 3mm and a t of < 15ms for all evaluation datasets. It suggested a good precision and real-time performance for PCA-NNE.