Feasibility of Using Pseudo-CT for Dosimetry, Radiomics, and Efficacy Assessment in IMRT/VMAT of Brain Tumors: A Multi-Omics Analysis

Abstract

Pseudo-CT generated by convolutional neural networks (CNN) and planning MRI has facilitated the promotion of MRI-Only. The technology not only reduces the time and money spent on CT scans, but also eliminates the cumbersome CT-MR registration. The feasibility in Stereotactic Brain Radiotherapy has been analyzed in previous studies by our team. However, when the prescribed requirements are not met, IMRT/VMAT are still selected. The study aims to evaluate the feasibility of pseudo-CT in IMRT/VMAT for brain cancer via the following 5 aspects: (1) image difference, (2) dose accuracy, (3) radiomics feature, (4) efficacy assessment, and (5) correlation analysis. Brain tumor patients who had received radiotherapy at our institution and had planning MRI and CT were included in the study. Redesign of IMRT and VMAT radiotherapy plans according to 3 × 15Gy for each patient. Hounsfield unit (HU) values for PTV and OARs were used to assess image differences. And dose accuracy analysis contained a 2D dose volume histogram (DVH) metrics (Dmax, Dmean, D2%, D50%, D98%, HI, CI) and 3D gamma metrics (criteria: 1-3%/2mm, 1%/1mm, 10% threshold). Then 107 original image features of PTV and OARs were extracted for radiometry analysis. And tumor control probability (TCP) of PTV (Poisson model) and normal tissue complication probability (NTCP) of OARs (Lyman-Kutcher-Burman model) were used for the variance analysis of efficacy assessment. Wilcox-test was used for significance of differences test (0.05), and spearman correlation analysis was used to explore the key features of the dose bias. A total of 42 patients were included, with 42 planning CTs and pseudo-CTs (mDixon-T1), and 38 pseudo-CTs (mDixon-T1-CE). The median volume of PTV was 4.1 cc (range 0.5-27.3), with no significant differences in HU, DVH, 3D gamma, and NTCP/TCP metrics. The median local gamma passing rates (1%/1mm) between planning CTs and pseudo-CTs (mDixon-T1) were 93.1% (range 65.5%-99.7%, IMRT) and 93.3% (range 63.9%-99.6%, VMAT). And more than 85% original radiomics features have significant difference. Further, the feature HU-Min was found to be more correlated with dose metrics in the correlation analysis. We speculate that it may be caused by the smoothing of the low frequency signal before outputting image. And since Shape_MeshVolume, Shape_VoxelVolume and PTV volume difference are highly correlated with dose deviation, it indicates that dose deviation affected by CT-MR registration. This study has the potential to provide guidance for the clinical application of pseudo-CT in the MRI-Only workflow with IMRT/VMAT for brain tumors. These quantitative results strongly indicate pseudo-CT can be used as a substitute for the initial CT in IMRT/VMAT for small brain lesions (size <5 cm, numbers <5), but not for radiomics analysis. Additionally, the impact of inter-image differences on dose accuracy is less significant compared to the deviation caused by image registration.