**1. Introduction**

In the treatment of abdominal tumors such as those of the liver and pancreas, carbon ion radiotherapy (CIRT) is considered a promising therapeutic option, thanks to its excellent geometrical selectivity and radiobiological effectiveness [1,2]. However, tumors subject to respiratory motion may suffer from inter- and intra-fraction motion that needs to be properly accounted for during planning and delivery. So far, the repeated acquisition of computed tomography (CT), in the form of respiratory-correlated 4DCT, is the clinical routine for motion management, but concurrently exposes the patient to additional nontherapeutic radiation [3].

The growing interest in magnetic resonance imaging (MRI) in recent years has fostered the development of MRI-only workflows that would guarantee the absence of ionizing radiation while exploiting the superior soft tissue contrasts of MRI scans. In this regard,

**Citation:** Parrella, G.; Vai, A.; Nakas, A.; Garau, N.; Meschini, G.; Camagni, F.; Molinelli, S.; Barcellini, A.; Pella, A.; Ciocca, M.; et al. Synthetic CT in Carbon Ion Radiotherapy of the Abdominal Site. *Bioengineering* **2023**, *10*, 250. https://doi.org/10.3390/ bioengineering10020250

Academic Editors: Paolo Zaffino and Maria Francesca Spadea

Received: 13 January 2023 Revised: 6 February 2023 Accepted: 9 February 2023 Published: 14 February 2023

**Copyright:** © 2023 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

online MR-guided radiotherapy (MRgRT) is a clinical reality in conventional radiotherapy [4–6], while no integrated system is clinically available for online MR guidance in particle therapy (PT), with only a few feasibility studies addressing protons but not carbon ions [7,8]. Nonetheless, the off-line use of MRI in CIRT may support treatment planning and adaptation through the generation of synthetic CT (sCT), used as alternatives to verification CTs, and avoid non-therapeutic radiation along with CT-MRI registration errors that are typically relevant in the abdominal site [9,10]. Different methods (i.e., bulk density override, atlas based, voxel based) have been studied to generate sCTs starting from MRI scans [8,11]. As an alternative, deep learning (DL) methods have been largely investigated [11–17], relying on their capability of autonomously learning hidden relationships among data [18]. In particular, deep convolutional neural networks (DCNN) showed promising results for the generation of sCTs of the head and neck with respect to conventional methods, as they are able to catch complex nonlinear relationships among MRI and CT [12,15]. Generative adversarial networks (GAN), especially in the conditional form (conditional GAN—cGAN) have been instead considered for more complex anatomical districts such as the abdomen, thanks to the adversarial learning process, which provides great efficacy in image-to-image translation applications [19–21]. The main limitation of these networks is the need for paired MRI-CT samples to perform the training, which is a critical aspect when dealing with multi-modal volumes. In recent years, new architectures have been proposed to overcome the need for paired CT-MRI training datasets, with cycleGAN and fully bidirectional networks showing great results in end-to-end synthetic image generation tasks [22]. Clearly, the use of such complex and multi-network architectures requires high dimensional training datasets, which still remain an evident concern in most clinical realities. The use of multiple MRI sequences (e.g., Dixon in-phase (IP), opposed-phase (OP), fat, and water) has been also investigated to increase the quality of critical anatomical structures, but no relevant improvements were shown [23].

The dosimetric evaluation of DL-based sCTs has been investigated for X-ray and proton radiotherapy in most anatomical districts [20–28], while a similar analysis for CIRT applications is still missing in the literature.

In the treatment of abdominal targets, the presence of inter- and intra-acquisition motions, which mainly translate into the different dispositions of the air fillings in the organs, highly affects the dose distribution for carbon ions plans [29]. An MRI-only workflow for CIRT would thus require more stringent constraints and tolerances in terms of accuracy in the definition of HU values, if compared to conventional radiotherapy, since a precise knowledge of particle stopping power, estimated from HU values within the patient anatomy, is essential for accurate treatment planning, to limit any range shift and damage to healthy tissues [30,31]. Up to now, to the authors' best knowledge, the application of sCT to CIRT plans has been investigated only by Knäusl et al. [32] for head and neck imaging, while no study has been performed on CIRT plans for abdominal tumors.

In this study, we focused on the generation of abdominal sCTs through a cGAN, with the final aim of simulating an MRI-only workflow for CIRT, based on recalculation of the clinical CIRT treatment plans for generated abdominal sCT volumes. This study, therefore, was conducted to evaluate, for the first time, the feasibility of sCT in CIRT for abdominal sites.

#### **2. Materials and Methods**

#### *2.1. Patient Cohort*

Image datasets were collected from 24 patients affected by liver or pancreatic cancer and treated with CIRT at the National Centre for Oncological Hadrontherapy (CNAO, Pavia, Italy) between 2017 and 2021. Standard clinical workflow comprised the acquisition of a 4DCT followed by a 3D MRI on the same day, for contouring and planning preparation. The same immobilization setup, consisting of customized pillows (MOLDCARE Cushion, QFix, Avondale, PA, USA) and non-perforated body thermoplastic masks (Klarity Medical Products, Heath, OH, USA), was used both for CT and MRI acquisitions and treatment

delivery. Acquisition in two different scanners along with re-positioning of the patient with the thermoplastic mask caused inter-acquisition motion. For 15 patients, re-evaluative images (both CT and MRI) were acquired during the treatment course with the same immobilizations setup and were considered independent from the first acquisition, leading to a total of 39 CT-MRI volume pairs collected. The study was approved by the local ethical committee, and all patients signed the informed consent (CNAO 37-2019 4D-MRI).

The 4DCTs were acquired during patient free breathing on a Siemens SOMATOM Sensation Open CT scanner (resolution 0.98 <sup>×</sup> 0.98 <sup>×</sup> 2 mm3). Clinical plans were optimized at end-exhale for gated treatments; as such, only this phase was used in this work to derive sCTs. CT acquisitions had a variable number of slices, resulting in a volume size of 512 × 512 × [96 − 145] voxels. MRI acquisitions were performed with a Siemens Magnetom Verio 3T scanner. Three-dimensional breath-hold T1-weighted volumetric interpolated breath-hold examination (VIBE) sequences were acquired at endexhale with 1.06 <sup>×</sup> 1.06 <sup>×</sup> 3 mm<sup>3</sup> resolution (repetition time TR = 3.87 ms, echo time TE = 1.92 ms). For two patients, MRI acquisitions had a voxel size of 1.25 <sup>×</sup> 1.25 <sup>×</sup> 3 mm<sup>3</sup> and 1.125 <sup>×</sup> 1.125 <sup>×</sup> 3 mm3. Most of the MRI acquisitions had 320 <sup>×</sup> <sup>260</sup> <sup>×</sup> 64 voxels, except for one having 88 transversal slices. Two CT-MRI volume pairs were discarded from the study because of the low quality of the acquired images. Therefore, 37 volume pairs were used: 32 pairs were exploited for cross-validation (CV) and training, while five pairs were randomly selected and held out for testing (Table 1) for more details on the treatment plans. All treatment plans were optimized with the RayStation (Raysearch Laboratories, Stockholm, Sweden—version 10.B) Treatment Planning System (TPS) on the end-exhale reconstructed CT phase and clinically approved. Corresponding organs at risk (OARs), gross tumor volume (GTV) and clinical target volume (CTV) were segmented by radiation oncologists. The relevant OARs included were kidneys, aorta, colon, duodenum, stomach and spinal cord.


**Table 1.** CIRT plan details for the patients used in the test set.
