Learning CT-to-electron-density mapping for radiotherapy treatment planning using neural networks

Electron density (ED) maps derived from spectral computed tomography (CT) offer superior accuracy—particularly in critical regions such as bone, metal implants, and areas with contrast agents—but their clinical adoption is limited by high equipment costs. This study aims to bridge this gap by developing a deep learning model that generates spectral-CT-comparable ED maps from conventional, widely available CT images. We propose a novel Dual Path Mamba Network (DPM-Net) framework. To address the challenges of precise image-to-image translation, DPM-Net leverages a Mamba-based architecture enhanced with residual and dense connection paths, effectively capturing long-range contextual dependencies crucial for modeling complex tissue densities. The model was trained and tested on a paired dataset of conventional CT images and their corresponding, high-fidelity spectral CT-based ED maps from 92 head-and-neck cancer patients. DPM-Net demonstrated state-of-the-art performance, outperforming models like UNet, UNetR, LightM-UNet, and UMamba by achieving reductions in Mean Absolute Error (MAE) of 0.1647, 0.2580, 0.0508, and 0.2384, respectively. Crucially, the dose distribution generated from our synthetic ED maps deviated by less than 0.5% from the dose plan based on the real spectral CT ED maps. The proposed DPM-Net framework enables accurate and clinically applicable electron density map generation from standard CT images for precision radiotherapy.