Learning-Based Stopping Power Mapping on Dual-Energy CT for Proton Radiation Therapy

Abstract

Purpose Dual-energy computed tomography (DECT) has been used to derive relative stopping power (RSP) maps by obtaining the energy dependence of photon interactions. The DECT-derived RSP maps could potentially be compromised by image noise levels and the severity of artifacts when using physics-based mapping techniques. This work presents a noise-robust learning-based method to predict RSP maps from DECT for proton radiation therapy. Materials and Methods The proposed method uses a residual attention cycle-consistent generative adversarial network to bring DECT-to-RSP mapping close to a 1-to-1 mapping by introducing an inverse RSP-to-DECT mapping. To evaluate the proposed method, we retrospectively investigated 20 head-and-neck cancer patients with DECT scans in proton radiation therapy simulation. Ground truth RSP values were assigned by calculation based on chemical compositions and acted as learning targets in the training process for DECT datasets; they were evaluated against results from the proposed method using a leave-one-out cross-validation strategy. Results The predicted RSP maps showed an average normalized mean square error of 2.83% across the whole body volume and an average mean error less than 3% in all volumes of interest. With additional simulated noise added in DECT datasets, the proposed method still maintained a comparable performance, while the physics-based stoichiometric method suffered degraded inaccuracy from increased noise level. The average differences from ground truth in dose volume histogram metrics for clinical target volumes were less than 0.2 Gy for D95% and Dmax with no statistical significance. Maximum difference in dose volume histogram metrics of organs at risk was around 1 Gy on average. Conclusion These results strongly indicate the high accuracy of RSP maps predicted by our machine-learning–based method and show its potential feasibility for proton treatment planning and dose calculation.

Machine learning based oxygen and carbon concentration derivation using dual‐energy CT for PET‐based dose verification in proton therapy

“…It is true that DECT images can be directly used to obtain C and O concentrations to generate the labels of the training dataset, through conventional approaches (e.g. pixel‐wise in the stoichiometric approach) 27 . Under such a circumstance, however, it would be quite challenging to ensure the fidelity of the dataset due to the inclusion of several intermediate steps, as well as the issue of CT number degeneracy (i.e., different elemental composition and density exhibiting the same CT number).…”

“…pixel-wise in the stoichiometric approach). 27 Under such a circumstance, however, it would be quite challenging to ensure the fidelity of the dataset due to the inclusion of several intermediate steps,as well as the issue of CT number degeneracy (i.e., different elemental composition and density exhibiting the same CT number). In the end, a pitfall of circular logic (e.g., deriving elemental composition using a non-machine learning approach as the training dataset for machine learning) might occur.…”

Deformable CT image registration via a dual feasible neural network

“…Compared with photon radiation therapy, proton radiation therapy features advantageous dosimetric properties attributed to the Bragg Peak and virtually no exit dose, which may have better clinical outcomes. [1][2][3][4][5] However, a small shift of the high dose gradient at the distal end of the Bragg peak can lead to substantial under-dose on target tumor or over-dose on critical organs. 6 Throughout the course of treatment, such shifts can be caused by patient setup uncertainty, tumor shrinkage, or patient weight loss, which results in delivered dose that deviates from the planned dose.…”

“…Compared with photon radiation therapy, proton radiation therapy features advantageous dosimetric properties attributed to the Bragg Peak and virtually no exit dose, which may have better clinical outcomes 1–5 . However, a small shift of the high dose gradient at the distal end of the Bragg peak can lead to substantial under‐dose on target tumor or over‐dose on critical organs 6 .…”

Generative Adversarial Network for Image Synthesis

“…In recent years, research in image synthesis gains great interest in radiation oncology, radiology and biology [85]. The presumed benefit has intrigued several investigations in a number of potential clinical applications such as magnetic resonance imaging (MRI)-only radiation therapy treatment planning [44,60,61,62,76,94], positron emission tomography (PET)/MRI scanning [90,36], proton stopping power estimation [9,10,35,86], synthetic image-aided auto-segmentation [16,20,29,49,39,53,63], low dose computerized tomography (CT) denoising [56,80,87], image quality enhancement [15,52,18,93], reconstruction [35,21], high resolution visualization [55] and etc. Historically, image synthesis methods have been investigated for decades.…”