Enhanced two-stage cascaded network for high-precision stopping power ratio and functional image generation from conventional CT.
Authors
Affiliations (7)
Affiliations (7)
- Department of Radiation Oncology, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China.
- School of Biomedical Engineering, Tsinghua University, Beijing, China; MedMind Technology Co., Ltd., Beijing, China.
- MedMind Technology Co., Ltd., Beijing, China.
- UMass Chan Medical School, Worcester, MA, USA.
- Medical Engineering Department, Peking Union Medical College Hospital, China.
- Department of Radiation Oncology, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China. Electronic address: [email protected].
- Department of Radiation Oncology, Peking Union Medical College Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China. Electronic address: [email protected].
Abstract
Accurate stopping power ratio (SPR) estimation is essential for proton therapy dose calculation; however, SPR and related functional images-electron density (ED), effective atomic number (Zeff), and iodine concentration (IC)-currently require dual-energy CT (DECT), whose clinical adoption is limited by high cost and restricted availability. To our knowledge, no previous study has synthesized DECT-derived functional images directly from conventional single-energy CT (SECT). This study proposes, for the first time, a deep learning framework that generates SPR, ED, Zeff, and IC maps simultaneously from conventional CT, and investigates whether a physics-informed two-stage cascaded architecture can outperform conventional single-stage networks. Seventy-four head-and-neck cancer patients scanned with a dual-layer DECT scanner were retrospectively analyzed. Integrated detector-layer signals provided SECT-equivalent input images, while DECT-derived maps served as references. Four networks were compared: a single-output UNet (UNet-1), a multi-output UNet (UNet-4), a two-stage cascaded UNet (Cascade-UNet), and the proposed enhanced two-stage cascaded model (Cascade-New), which incorporates an asymmetric encoder-decoder with residual vision modules and Mamba-based dense connectivity, and exploits the physical dependency chain CT → ED/Zeff → SPR through a three-term training loss. Models were evaluated on an independent test set using mean absolute error (MAE), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and Pearson correlation coefficient (PCC), with pairwise Wilcoxon signed-rank tests. The proposed Cascade-New model achieved the best overall performance: the lowest SPR MAE (11.65 ± 1.58 in units of SPR × 10<sup>-3</sup> vs. 12.99 ± 1.46 for UNet-4, a 10.3% relative reduction), the highest SSIM (0.9971 ± 0.0016), and the strongest correlation (PCC = 0.9996 ± 0.0002). Improvements over UNet-4 in ED, Zeff, and IC synthesis were all statistically significant (all p = 0.031). Visual and error map analyses confirmed its improved performance, particularly in bone and heterogeneous tissue regions. This study provides the first demonstration that SPR and other DECT-derived functional images can be synthesized directly from conventional SECT with high precision. The proposed physics-informed cascaded framework offers a cost-effective alternative to DECT, enabling radiotherapy centers to obtain accurate SPR maps with conventional CT for proton dose calculation, potentially reducing range uncertainty and expanding access to high-precision functional imaging.