A machine learning based dual-energy CT elemental decomposition method and its physical-biological impacts on carbon ion therapy

Background

Dual-energy computed tomography (DECT) enhances material differentiation by leveraging energy-dependent attenuation properties particularly for carbon ion therapy. Accurate estimation of tissue elemental composition via DECT can improve quantification of physical and biological doses.

Objective

This study proposed a novel machine-learning-based DECT (ML-DECT) method to predict the physical density and mass ratios of H, C, N, O, P, and Ca. The physical and biological impacts on carbon ion therapy was also investigated.

Methods

Taking DECT-derived CT numbers as inputs, a fully connected neural network was employed to predict the physical density or the elemental mass ratio. The training and testing utilized a dataset of 85 biological tissues with data augmentation. The prediction accuracy and noise analysis were compared against the parameterization DECT (PA-DECT) and SECT methods. By applying the proposed method on the DECT images of 10 head-and-neck patients, the physical and biological doses as well as the linear energy transfer (LET) were calculated for a set of carbon ion pencil beams using Monte-Carlo simulations. Patient-based results were compared with the PA-DECT method.

Results

The ML-DECT method yielded $R^2=0.9996$ for physical density and $R^2=0.8338\sim 0.9997$ for the six elemental mass ratios across 85 materials. Compared to the PA-DECT and SECT methods, the accuracy was improved by over 20% and 50%; the noise robustness was improved by over three times and up to 25%, respectively. In the patient dose evaluation, the ML-DECT method yielded comparable physical and biological doses, yet up to ∼1% higher LET, and up to ∼2 mm shallower peak positions than those of the PA-DECT method.

Conclusion

The ML-DECT method provided precise estimation of physical density and elemental mass ratios of human tissues. Compared with the PA-DECT method, the ML-DECT method displayed stronger robustness to image noise.