Estimation of proton stopping power from principal components of dual‐energy CT virtual monoenergetic images

Background DECT systems can perform projection‐domain decomposition to generate quantitative maps of electron density () and effective atomic number () for proton stopping power ratio (SPR) estimation. In contrast, image‐based methods require reconstructed images and, when low‐ and high‐energy polyenergetic images are unavailable, rely on selected pairs of virtual monoenergetic images (VMIs). However, fixed VMI pairs ignore information contained in the full VMI set; a data‐driven combination of VMIs can better exploit this information and improve and SPR estimation. Purpose To develop and evaluate a principal component analysis (PCA)‐based method that optimally combines the full set of VMIs into a single image pair for estimating and SPR, and to validate its performance using both phantom experiments and simulated patient datasets. Methods PCA is applied to CT numbers of tissue‐equivalent inserts across VMIs reconstructed at 50–200 keV to derive two independent component images, which are used as input to SPR estimation frameworks. Performance is evaluated on a tissue‐equivalent phantom using leave‐one‐out cross‐validation and further assessed on realistic simulated patient CT datasets. Results are compared with conventional VMI pairs, polyenergetic images, and single‐energy CT (SECT) methods. Results In the phantom study, the PCA‐based method reduced RMS errors in and SPR to 0.94% and 1.02%, compared with 1.08% and 1.19% for estimation from the 50/200 keV VMI pair, and 1.36% and 2.45% for scanner‐derived outputs. In simulated patient datasets, PCA‐derived VMIs consistently achieved the lowest RMS SPR errors across both DECT and SECT scenarios, outperforming alternative image inputs. Conclusions The proposed PCA‐based strategy provides a VMI input pair by combining all available VMIs along principal component directions, producing images that can be used for accurate and SPR estimation. Consistent improvements in both phantom and simulated patient studies demonstrate its generalizability and potential for proton therapy treatment planning.