The landscape of artificial intelligence in carbon ion radiotherapy: A scoping review

Abstract

Objective:

While artificial intelligence (AI) is known to benefit radiotherapy (RT), its applications in carbon ion RT (CIRT) is yet to be surveyed. This scoping review aims to examine the current landscape of AI in CIRT and highlight future directions.

Materials and methods:

Following PRISMA guidelines, three databases (PubMed, Scopus, and Web of Science) were searched to find eligible English articles investigating AI in CIRT between January 2014 and June 2025. Two human reviewers and an AI tool independently screened articles. Agreement was quantified using Cohen’s kappa and McNemar’s test (α = 0.05).

Results:

Of the 413 unique articles screened, 16 were selected, and an additional three were included through citation search, totaling 19 articles. The screening agreement between human and AI reviewers was substantial (Cohen’s kappa = 0.75) with no statistical difference (p = 0.18). The AI contributions in CIRT were categorized into treatment planning, optimization, and verification; synthetic imaging; tumor control; and normal tissue complication prediction. All selected studies were limited by small datasets and lacked external validation.

Conclusions:

AI development in the CIRT domain is in early stages but holds promise for reducing costs and improving treatment, especially in adaptive settings. Future efforts should foster collaboration across CIRT facilities to support robust AI applications.

Keywords

carbon ion radiotherapy artificial intelligence medical informatics applications large language models synthetic imaging adaptive therapy

Introduction

In recent years, an increasing interest in artificial intelligence (AI) approaches has been reported in healthcare.¹ This is largely driven by significant technological advancements over the past decade, including hardware innovations and enhanced accessibility of key hardware components required to fully leverage powerful AI algorithms, such as deep learning (DL). The contextual data digitalization in healthcare further facilitated the proliferation of AI applications. Radiotherapy (RT) was one of the most explored fields, likely due to the large availability of multimodal imaging datasets, especially after digitalization. Based on a recent review,² most of the AI efforts in photon RT are mainly focused on treatment plan optimization, image quality improvement, and treatment outcome prediction. A similar pattern was also observed in proton therapy (PT),³ although with lower contributions due to the limited widespread use of this treatment modality. Notably, AI applications have been progressively developed and are starting to be integrated into routine clinical practice to increase RT efficiency. A key example is the automatic contouring of organs of interest,⁴ which is now commercially available in most treatment planning systems (TPS) and is generally powered by AI.

Carbon ion radiotherapy (CIRT), an advanced form of RT, leverages carbon ions’ peculiarities to gain geometrical and radiobiological advantages in the treatment of deep-seated, unfavorably located, and radioresistant tumors. CIRT offers highly favorable treatment outcomes, albeit with substantial technological costs. Only a few CIRT facilities are operating worldwide, hence the availability of such a promising treatment is currently limited.

Substantial economic advantages in many healthcare fields were reported when integrating AI in key clinical⁵ and healthcare administrative processes.⁶ In this scenario, AI may represent a pivotal tool to mitigate CIRT treatment costs and increase its accessibility. Furthermore, to promote economically convenient resource allocation, AI tools may serve to effectively select patients who would likely gain the highest benefit from CIRT.⁷ Beyond the economic aspects, CIRT could also gain significant advantages from AI-guided personalized treatments.⁸ Finally, due to the peculiar characteristics of carbon ion interaction with biological tissues, CIRT is particularly sensitive to inter-fraction anatomical modifications, and adaptive treatments are warranted to maximize CIRT outcomes. For such adaptive settings, AI-based frameworks are particularly well-suited to address the stringent demands of both accuracy and computational efficiency.

This scoping review aims to systematically examine the current landscape and status of AI applications in CIRT, while critically identifying gaps and future directions.

Materials and methods

This scoping review was conducted following the PRISMA Extension for Scoping Reviews (PRISMA—ScR)⁹ to describe the current status of AI applications in the landscape of CIRT. The PRISMA-ScR checklist was filled out and is provided as Online Supplementary Material (SM).

Eligibility criteria

The population, concept, and context framework was used to define the eligibility criteria.¹⁰ No limitations regarding population were defined, and articles on AI (concept) applications in CIRT (context) were included. Specifically, scientific manuscripts exploring, developing, or taking advantage of any AI method in the context of CIRT were included in the literature analysis. Articles not written in English were excluded.

Information sources and search

Manuscripts published between January 2014 and June 2025 were retrieved from the following three databases: Web of Science, Scopus, and PubMed. The search string (see SM, S1) was constructed to include the following keywords within the manuscript’s title or abstract:

• carbon ion radiotherapy, carbon ion radiation therapy, carbon ion RT, CIRT, particle therapy, hadron therapy, heavy ion radiotherapy, heavy ion radiation therapy, heavy ion RT; AND

• artificial intelligence, AI, machine learning, deep learning, neural network, model, radiomics, dosiomics, omics, bioinformatics, sequencing, LLM, NLP, auto-segmentation, auto-contouring, GANN, CNN, large language model, natural language processing, synthetic, prediction, predictive.

Relevant scoping, narrative, or mapping reviews exploring AI in the field of particle RT were evaluated through citation search to account for potentially pertinent articles not captured by our search string.

The following metadata were collected from all sourced manuscripts: authors, title, date of publication, volume, issue, journal, publisher, DOI, abstract, URL, keywords, and location.

Selection of sources of evidence

Metadata from all the retrieved manuscripts were imported into Rayyan¹¹ for the screening process. Duplicate entries were automatically detected and were deleted manually after inspection. Two human reviewers and a fully open-source AI large language model (LLM) reviewer (DeepSeek-R1:32B)¹² independently screened the unique manuscripts based on the article title and abstract. The AI reviewer was prompted to provide a simple yes or no answer along with a concise one-sentence description for the decision (see SM, S2 for prompt details). Only human decisions were considered for source selection purposes, and disagreements among them were resolved by consensus.

Cohen’s kappa¹³ was used to measure the screening agreement between consensual human and AI reviewer decisions. Metrics such as sensitivity, specificity, and precision were also measured¹³. Subsequently, McNemar’s test was applied to determine if there was a statistically significant difference in article selections between the human and AI reviewers. The significance level was set to 0.05. Finally, studies not identified with the search string, but found pertaining to the review scope through citation search, were also included in the analysis.

Data charting process and data items

The two human reviewers evaluated the selected manuscripts and classified them into different AI application categories within the CIRT domain. A spreadsheet was designed to collect the following data items from each included article: application category, title, authors, journal, year of publication, aim, dataset, dataset modality, site, AI methodology, AI performance metrics, and LLM summary. For tabulation, data were primarily extracted manually from the articles’ abstract, while the full text was consulted only when deriving the data items was not possible from the abstract alone. One human reviewer independently extracted the data, while the second human reviewer subsequently verified the entries. The LLM summary of the selected manuscripts was obtained by asking the AI reviewer to respond to a specific set of tailored questions and subsequently combine its responses to generate a short summary (see section SM, S3 for prompt details), using the full manuscript text (excluding the references sub-section to reduce the context length). The generated AI summaries were included after careful assessment of their correctness by one of the human reviewers, as LLM models are prone to hallucinations.

Synthesis of results

Word clouds were used to characterize the AI terminology and identify trends in the CIRT research domain. A comprehensive critical summary for each defined AI category was elaborated by one of the two human reviewers. All the exploratory trends were plotted using Seaborn-0.13.2¹⁴ and wordcloud-1.9.4¹⁵ packages, while the AI-based abstract screening and critical summary pipelines were implemented using Ollama-0.6.2¹⁶ and LangChain-0.3.20,¹⁷ all in Python-3.13.0. To support future reviews leveraging fully open-source locally run LLM for screening and critical review, the code was made open source and is available at: https://github.com/sithin-cnao/llm_review.git (accessed on 22 January 2026).

Results

A total of 689 records were retrieved from the three databases. After removing duplicates, 413 unique articles were screened. Of these, 16 were included in the study after consensus between the two human reviewers. The PRISMA flowchart¹⁸ is reported in Figure 1. On the other hand, the AI reviewer selected 21 articles, 14 of which overlapped with the human reviewers’ selection. When quantitatively compared against the consensual human selections, the AI reviewer demonstrated 88% sensitivity and 98% specificity, but with a moderate precision of 67%. The agreement between the human reviewers and the AI reviewer was substantial, with a Cohen’s kappa of 0.75 and no statistically significant difference in agreement (p = 0.18).

Figure 1.

PRISMA flowchart describing the source of evidence retrieval and selection process of the scientific literature on artificial intelligence applications in the context of carbon ion radiotherapy. The experimental LLM screening was excluded from the flowchart, as the final inclusion decisions relied exclusively on humans.

The AI methods associated with abstracts where human and AI reviewers’ decisions matched and mismatched were identified through word clouds, and are reported in Figure 2. The terms related to heuristic search (eg., A*), scheduling algorithms, and LASSO were found to be mainly related to the AI reviewer’s incorrect manuscript selection. This highlighted the LLM’s inclination to classify conventional statistical tools or heuristic AI optimization algorithms as AI applications.

Figure 2.

Word cloud of AI methodologies extracted from the screened abstracts where human and LLM-reviewer agreed (black) and disagreed (red).

Figure 3 presents the AI methods explored in the selected articles according to their year of publication. Interestingly, generative AI methods were mostly represented in the period from 2022 onwards, while conventional machine learning (ML) approaches were more common in earlier years. Excluding AI-unrelated terms, the most common AI methods explored in the context of CIRT were: generative adversarial networks (GAN), radiomics, U-Net, and ML methods such as logistic regression, support vector machines (SVM), random forest, etc. Word clouds were used to detect such trends (see Figure 3b).

Figure 3.

General trend of AI literature in CIRT. (a) Number of publications per year. (b) Word clouds of AI methods explored in the selected articles, grouped by year.

The CIRT AI applications investigated in the selected articles were classified into four categories: (i) Treatment planning, optimization, and verification (# articles = five, publication years = 2020-2025), (ii) Synthetic imaging (five, 2022-2025), (iii) Tumor control prediction (five, 2019-2024) and finally (iv) Normal tissue complication prediction (four, 2018-2025). Table 1 presents the manually extracted data items alongside the human-revised LLM-generated summary, while the LLM’s response to critical review questions for the selected papers is provided in the Online Supplementary Material. A comprehensive overview of each category is presented in the subsequent paragraphs.

Table 1.

Manually extracted data items and LLM-generated short summaries of 19 selected articles.

Application category	Title	Authors	Journal	Year of publication	Aim	Dataset	Dataset modality	Site	AI methodology	AI performance metrics	LLM Summary
Synthetic imaging	DR-only Carbon-ion radiotherapy treatment planning via deep learning	Zhang et al	Physica Medica	2022	Feasibility of patient-specific DR-only treatment planning	Anthropomorphic thorax-abdomen phantom; 2 Head and Neck Cancer (HNC) patients	X-ray DR, CBCT, CT	Thorax, Abdomen ,Head and Neck	cycle GAN;Custom deep neural network (encoder - transformation - decoder modules)	Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM); 3D gamma analysis on actual predicted dose differences	The study demonstrates the feasibility of using deep learning to generate CT images from DR for CIRT, achieving high accuracy with potential for adaptive treatment planning.
Synthetic imaging	Synthetic CT in Carbon Ion Radiotherapy of the Abdominal Site	Parrella et al	Bioengineering	2023	Feasibility of sCT in CIRT for abdominal sites.	24 pancreatic/liver cancer patients 4D-XCAT CT / ComBat MRI phantom	MRI, CT	Abdomen	Three channels 2D conditional GAN (cGAN): U-Net generator and PatchGAN discriminator	Motion analysis, similarity metrics [e.g. MAE, Normalized Cross Coefficient (NCC)]; Geometric and dosimetric criteria	The study evaluates a cGAN model for generating sCT volumes in CIRT, demonstrating acceptable accuracy with MAEs of 57.08 HU (body), 86.03 HU (bone), and 55.39 HU (soft tissue). While promising, challenges remain in accurately representing air pockets and bone structures.
Synthetic imaging	Possibilities and challenges when using synthetic computed tomography in an adaptive carbon-ion treatment workflow	Knausl et al	Zeitschrift für Medizinische Physik	2023	To evaluate the possibilities and challenges of using MRI-based sCT in CIRT	30 HNC patients	MRI, CT	Head and Neck	3D U-Net	Dose differences and carbon-ion spot displacements between sCT and X-rayCT	The study evaluates sCT from MRI in CIRT, finding promise but highlighting challenges like inaccuracies in bone/air cavities and workflow limitations. It concludes sCTs are not yet ready to replace X-rayCT.
Synthetic imaging	Unsupervised Deep Learning for Synthetic CT Generation from CBCT Images for Proton and Carbon Ion Therapy for Paediatric Patients	Pepa et al	Sensors	2024	To evaluate a cycle GAN-based sCT generation from CBCT towards adaptive Particle therapy in paediatric patients	21 paediatric and young adult patients	CBCT, CT	Pelvis ,Abdomen	Cycle GAN	SSIM, and PSNR	The article explores the use of CycleGAN to generate sCT scans from CBCT images for adaptive particle therapy, demonstrating improved image quality and potential clinical benefits in reducing radiation exposure and enhancing treatment accuracy.
Synthetic imaging	Deep-learning synthetized 4DCT from 4DMRI of the abdominal site in carbon-ion radiotherapy	Nakas et al	Physica Medica	2025	Feasibility of deep-learning-based synthetic 4DCT (4D-sCT) generation from 4D-MRI data of abdominal patients	26 pancreatic/liver cancer patients ;4D-XCAT CT / ComBat MRI phantom	MRI, CT	Abdomen	Three-channel 2D cGAN: U-Net generator and PatchGAN discriminator	Motion analysis, similarity metrics [e.g. Mean Absolute Error (MAE), Normalized Cross Coefficient (NCC)), motion analysis and dosimetric criteria	The study uses a DL model to generate 4D-sCT from 4D-MRI for CIRT, achieving accurate image synthesis with good dosimetric consistency. While challenges remain in air channel accuracy and OAR dose discrepancies, the approach shows promise for reducing radiation exposure and enhancing treatment planning efficiency.
Tumor control prediction	Imaging-Based Individualized Response Prediction Of Carbon Ion Radiotherapy For Prostate Cancer Patients	Wu et al	Cancer Management and Research	2019	To explore the value of baseline MRI radiomic features in predicting response to CIRT for prostate cancer patients	23 prostate cancer patients	MRI (T2-weighted, Apparent Diffusion Coefficient)	Pelvis	Radiomics - Support Vector Machine (SVM)	Accuracy, sensitivity, specificity, Area Under the Receiver Operating Characteristic Curve (AUROC); Wilcoxon test.	The study uses radiomic features from MRI to predict CIRT response in prostate cancer patients, achieving high accuracy with SVM despite a small sample size.
Tumor control prediction	A Comparison Study of Machine Learning (Random Survival Forest) and Classic Statistic (Cox Proportional Hazards) for Predicting Progression in High-Grade Glioma after Proton and Carbon Ion Radiotherapy	Qiu et al	Frontiers in Oncology	2020	To compare Random Survival Forest (RSF) and Cox Proportional Hazard (CPH) in predicting tumor progression of High-Grade Glioma (HGG) after particle therapy	82 HGG patients	Patient-, tumor- and treatment-related variables	Brain	RSF, CPH	Concordance index (C-index) for discrimination (accuracy), Brier score for calibration (precision) and variable importance for interpretability	The study compares RSF and CPH models for predicting tumor progression in HGG patients treated with proton and carbon ion radiotherapy. With a dataset of 82 patients, CPH outperformed RSF in discrimination and calibration, emphasizing the need for comprehensive evaluation criteria when deploying AI models clinically.
Tumor control prediction	Radiomics and Dosiomics for Predicting Local Control after Carbon-Ion Radiotherapy in Skull-Base Chordoma	Buizza et al	Cancers	2020	To explore radiomic approaches for predicting local control in skull-base chordoma treated with CIRT	57 skull-base chordoma patients	MRI (T1-weighted, T2-weighted), CT, dose maps	Skull-base	Radiomics/Dosiomics - survival-SVM, regularized-CPH	C-index, Log-rank test after model classification	The study explored AI applications in predicting local control after CIRT using radiomic, dosiomic, and clinical features. Dosiomic models performed best, but small sample size limits generalizability. Further validation is needed.
Tumor control prediction	A Dosiomics Analysis Based on Linear Energy Transfer and Biological Dose Maps to Predict Local Recurrence in Sacral Chordomas after Carbon-Ion Radiotherapy	Morelli et al	Cancers	2023	To compare the performances of survival models using dosiomics features and conventional Dose-Volume Histogram (DVH) metrics extracted from relative biological effectiveness (RBE)-weighted dose (DRBE) and dose-averaged Linear Energy Transfer (LETd) maps	50 sacral chordoma patients	DVH parameters, DRBE and LETd	Pelvis	Dosiomics/LET-omics survival-SVM, regularized-CPH	C-index, Log-rank test after model classification	The study applies AI to predict local recurrence in sacral chordomas using biological dose maps and LETd, showing that dosiomics outperforms traditional metrics. However, the findings are limited by a small dataset and need further validation.
Tumor control prediction	A dosiomics approach to treatment outcome modeling in carbon ion radiotherapy for skull base chordomas	Parrella et al	Physica Medica	2024	To investigate dosiomics features from physical dose, DRBE and LETd, to predict local recurrence in skull base chordoma treated with CIRT; To define and evaluate dosiomics-driven Tumor Control Probability (TCP) models	54 skull-base chordoma patients	CT, dose maps, LETd maps	Skull-base	Dosiomics/LET-omics survival-SVM, regularized-CPH Logistic regression-based TCP models	C-index, Log-rank test after model classification. F-score and accuracy on TCP models	The study investigates AI applications in CIRT by developing dosiomics-based predictive models for local recurrence in skull base chordomas. Using a dataset of 54 patients, the authors employed regularized CPH models and survival-SVM to analyze features from physical dose, DRBE and LETd maps. The results indicate that LETd-based models performed best. Additionally, the dosiomics-driven TCP models demonstrated superior performance compared to clinical models.
Normal tissue complication prediction	Prediction model of critical weight loss in cancer patients during particle therapy	Zhang et al	Japanese Journal of Clinical Oncology	2018	To build a predictive model for critical weight loss in cancer patients receiving particle therapy	365 cancer patients	Patients’ demographic, clinical characteristics, pre-therapeutic nutrition risk screening and BMI	N.A.	Logistic regression, Chi squared Automatic Interaction Detection decision tree analysis (CHAID)	Accuracy, AUROC	The study predicts critical weight loss in particle therapy using logistic regression and decision trees, identifying key predictors with good accuracy but noting limitations like small sample size.
Normal tissue complication prediction	Normal Tissue Complication Probability Model for Acute Radiation Dermatitis in Patients With Head and Neck Cancer Treated With Carbon Ion Radiation Therapy	Li et al	International Journal of Radiation Oncology, Biology, Physics	2022	To build a Normal Tissue Complication Probability (NTCP) model for acute radiation dermatitis in patients with HNC treated with CIRT	187 HNC patients	patient-specific and treatment-related factors	Head and Neck	Logistic regression	10-fold cross-validated AUROC, Calibration plots	The study developed an NTCP model using 187 HNC patients treated with CIRT to predict Acute Radiation Dermatitis (ARD) incidence. Multivariate logistic regression analyzed biological and physical dose parameters, showing comparable performance between them. Further validation is needed for broader applicability.
Normal tissue complication prediction	Evaluation of machine learning models for predicting xerostomia in adults with head and neck cancer during proton and heavy ion radiotherapy	Zhang et al	Radiotherapy and Oncology	2025	To evaluate the performance of machine learning approaches in predicting Grade 2 or higher (G2+) xerostomia in adults with HNC receiving proton and carbon ion radiotherapy	1769 HNC patients	35 variables (e.g demographic, nutrition status, tumor-related and treatment-related factors, laboratory test results)	Head and Neck	Eight machine learning models (e.g. linear-SVM, logistic regression model)	Balanced accuracy for model selection, AUROC, precision, recall, and F1 scores	The study assessed AI models' performance in predicting xerostomia during CIRTHNC patients. Despite challenges with imbalanced data, an SVM model achieved moderate success, emphasizing the need for advanced predictive methods.
Normal tissue complication prediction	Normal tissue complication probability model for acute oral mucositis in patients with head and neck cancer undergoing carbon ion radiation therapy based on dosimetry, radiomics, and dosiomics	Meng et al	Radiotherapy and Oncology	2025	To develop a NTCP model for predicting G2+ acute oral mucositis in HNC patients undergoing CIRT	178 HNC patients	clinical, DVH, radiomics, and dosiomics data	Head and Neck	Logistic regression	AUROC and the corrected Akaike Information Criteria (AIC), bootstrap calibration plot, and confusion matrices	The study developed a hybrid NTCP model for predicting acute oral mucositis in CIRT patients using DVH, radiomics, and dosiomics data. It achieved high performance (AUC: 0.932 training, 0.959 test) and outperformed single-modality models, aiding clinical decision-making.
Treatment planning, optimization, and verification	Prediction of dose distribution from luminescence image of water using a deep convolutional neural network for particle therapy	Yabe et al	Medical Physics	2020	To estimate dose distributions from the measured luminescence images of water using a Deep Convolutional Neural Network (DCNN)	water phantom	luminescence images in water, dose distributions	N.A.	U-Net	Root Mean Square Percentage Error (RMSPE) and SSIM	The study uses a U-Net model to predict 2D dose distributions from luminescence images in proton and carbon ion therapy, achieving higher accuracy than conventional methods. It demonstrates improved performance metrics and faster predictions, though some limitations remain with real-world data.
Treatment planning, optimization, and verification	Deep learning-based fast denoising of Monte Carlo dose calculation in carbon ion radiotherapy	Zhang et al	Medical Physics	2023	To evaluate the feasibility of using a deep learning-based Monte Carlo (MC) denoising method to accelerate CIRT plan verification	59 patients (29 HNC and 30 lung cancer patients)	dose distributions	Head and Neck; Lung	CycleGAN, 3DU-Net and GhostU-Net	RMSE and 3D gamma analysis (3 mm, 3%).	The study evaluates deep learning methods to accelerate plan verification in CIRT by denoising MC simulations. Using CycleGAN, 3DU-Net, and GhostU-Net models, the research demonstrates that GhostU-Net achieves superior performance with fewer parameters and faster computation.
Treatment planning, optimization, and verification	Deep learning-based voxel sampling for particle therapy treatment planning	Quarz et al	Physics in Medicine & Biology	2024	To present a novel Deep-Learning (DL) model to select a subset of voxels in the planning process toward computational efficiency improvement	70 HNC patients	Treatment planning data	Head and Neck; Skull-base	P-Net	Dosimetric criteria, gamma analysis	The study introduces a DL model that reduces voxel sampling in CIRT, cutting computation time by 84.8% with minimal plan quality loss. It effectively handles larger tumors but shows trade-offs for smaller ones, demonstrating AI's potential to improve treatment planning efficiency.
Treatment planning, optimization, and verification	Deep learning-based Monte Carlo dose prediction for heavy-ion online adaptive radiotherapy and fast quality assurance: A feasibility study	He et al	Medical Physics	2025	To develop a DL model for independent MC dose prediction, aiming to facilitate online adaptive RT and rapid quality assurance	120 HNC patients	patient CT,3D energy matrices, and ray-masks	Head and Neck	Channel Attention Mechanism (CAM)-CHD U-Net	Dosimetric criteria, gamma analysis	The study presents a DL model (CAM-CHD U-Net) for predicting MC doses in heavy-ion therapy, demonstrating high accuracy with 99.31% gamma passing rate in PTV but noting limitations like dataset diversity and lack of RBE consideration.
Treatment planning, optimization, and verification	Dose image prediction for range and width verifications from carbon ion induced secondary electron bremsstrahlung x-rays using deep learning workflow	Yamaguchi et al	Medical Physics	2020	To predict the dose images for estimating the range and width of the carbon ion beam on the measured Secondary Electron Bremsstrahlung (SEB) x-ray images	10000 dose/SEB image pairs; 500 simulated/measured SEB	x-ray, dose images	N.A.	U-Net	average mean squared error (MSE), average SSIM, beam range and width error distances	The study uses deep learning to predict dose images from SEB x-rays in CIRT, achieving high accuracy in simulations and moderate results in real measurements, suggesting further clinical validation is needed.

Treatment planning, optimization, and verification

Five studies were published between 2020 and 2025, encompassing AI applications related to treatment planning or dose distribution prediction in CIRT. The selected papers showed heterogeneity not only in AI methodologies but also in their application scopes and have not undergone external validation. Nonetheless, they mostly aimed to overcome the resource and time consumption limitations of Monte Carlo (MC) simulation approaches toward adaptive CIRT, predominantly using DL based on U-Net⁴ architecture.

In the context of particle therapy range verification, in 2020, Yabe et al.¹⁹ successfully developed a U-Net model to predict accurate and reliable dose distributions almost instantaneously (<1s) from measured luminescence water images during proton or carbon ion treatment. This was an extension to their previous work on protons.²⁰ Similarly, Yamaguchi et al.²¹ proposed a U-Net model to predict the dose images with low error margins from secondary bremsstrahlung electron (SEB) X-ray images. On the other hand, Zhang et al.²² focused on plan verification and quality assurance. In this context, conventional phantom-based methods are laborious and time-consuming, while MC approaches require substantial resources and time.²³ The authors explored three DL approaches (cycle-GAN²⁴, 3D U-Net²⁵, Ghost-U-Net²⁶) for MC-dose simulation denoising in lung and head and neck cancer patients undergoing CIRT. This approach was previously explored and was oriented to preserve the MC simulations’ dosimetric accuracy while reducing the particles used for the simulation^27,28. Along the same scope, He et al.²⁹ developed a custom U-Net model for end-to-end dose prediction in online head and neck adaptive CIRT. Unlike Zhang et al, the authors did not include MC simulations in their proposed pipeline; instead, doses were predicted directly from the treatment planning computed tomography (CT), the 3D energy matrices, and the ray masks. Similarly, to facilitate adaptive planning, Quarz et al.³⁰ aimed to mitigate MC time consumption by leveraging DL models for voxel sampling in CIRT. While heuristic approaches in defining the relevant voxels were investigated, such methods remain patient-specific and are still labor-intensive and time-consuming.³¹ Hence, the authors explored the feasibility of a novel AI model, P-Net,³² in deriving the optimal sampling rate within the target and organs of interest volume regions, which was found to generate high-quality plans while reducing computational burden.

Synthetic imaging

Five studies published between 2022 and 2025 explored generative AI imaging applications in the domain of CIRT. All of them focused on generating synthetic CT (sCT) from X-ray digital radiography (DR), magnetic resonance imaging (MRI), or cone beam CT (CBCT). The main benefits identified were the potential integration of such tools in an adaptive CIRT framework and/or an MRI-only treatment planning.

In July 2022, Zhang et al.³³ published the first paper on sCT in CIRT, aiming to evaluate the feasibility of generating sCT from a single setup verification DR. The authors built a two-step framework where an improved cycle-GAN and a customized deep neural network models were trained to map DR to digitally reconstructed radiography (DRR), and DRR to CT, respectively. Despite performance metrics and gamma pass-rate being acceptable, the dataset was extremely small, and consisted only of a thorax-abdomen anthropomorphic phantom and two head and neck patients. Furthermore, concerning the clinical cases, the reference treatment planning was only optimized on the treatment positioning CBCT, further limiting their conclusions. Knäusl et al.³⁴ explored a more popular research approach and evaluated the feasibility of using MRI to generate sCT using a 3D U-Net model in head and neck CIRT. Nonetheless, the relevant dosimetric differences between sCT- and CT-based treatment plans highlighted the need for further improvements in the methodology before such an approach could be integrated into an adaptive CIRT framework. The treatment plans’ dosimetric differences were mainly ascribed to the immobilization mask being poorly reconstructed in the sCT, as well as the target and organs of interest volume-specific positions. Parrella et al.³⁵ and Nakas et al.³⁶ investigated the feasibility of generating sCT and four-dimensional sCT (4D-sCT), respectively, from virtual 4D-MRI to mitigate the increased exposure to ionizing radiation during long CT procedures and potential errors due to abdominal breathing motion. In both cases, three-channel conditional-GANs³⁷ were built, demonstrating good performance and acceptable treatment plans compared to the 4D-CT reference plans. Overall, both Parrella et al. and Nakas et al. results were aligned with the literature investigating sCT in the abdominal site. Nonetheless, in the context of an MRI-only scenario, both reported a few cases with sub-optimal results, likely due to air pockets and inter-acquisition variability as reported in previous literature studies on proton RT.³⁸ Finally, Pepa et al.³⁹ explored the feasibility of using cycle-GAN to generate sCT from CBCT acquired for patient positioning in pediatric CIRT. Although the authors reported good results, their evaluation was only based on imaging, with no accompanying dosimetric analysis. On the other hand, Klassen et al.⁴⁰ conducted a dosimetric-oriented synthetic study to evaluate the potential advantages of sCT from CBCT for hypo-fractionated adaptive pancreatic CIRT.

Tumour control prediction

Five manuscripts published between 2019 and 2024 investigated the potential of AI in tumor control prediction. Most of these manuscripts (four out of five) were based on a radiomics approach or its derivatives. All of the papers took advantage of simple conventional ML methods rather than complex DL models. Note that none of the AI models developed in these studies has undergone external validation.

Wu et al.⁴¹ investigated MRI radiomics for both T2-weighted and apparent diffusion coefficient maps to predict prostate tumor relapse after CIRT. While the robustness of the radiomics features was considered, the opportunity to properly evaluate the model was strongly limited by an extremely reduced sample size (n = 23). A support vector machine (SVM) model was proposed, and its performance evaluation was carried out with leave-one-out cross-validation area under the receiver operating characteristic curve (AUROC). Despite the high AUROC (0.88), no further studies were conducted to investigate the validity of their findings. In 2020, Buizza et al.⁴² investigated the potential benefit of multimodal (MR, CT) radiomics, dosiomics (using 3D dose maps extracted from the treatment plan), and clinical features in predicting local relapse of skull-base chordoma patients after CIRT. Unlike Wu et al, the authors chose to implement a survival-SVM (s-SVM) over its classification variant (SVM) to stratify patients based on their risk for adverse outcomes. Such a decision was likely supported by the higher interest in the survival outcome rather than in the mere prediction of relapse, which is quite common in radiotherapy. Additionally, a comparison with more common statistical approaches was conducted by building a regularized Cox-proportional hazard (r-Cox) model. Their best model capable of effectively separating high and low risk groups was dosiomics-based (shape and texture features), while the lowest generalizability was observed for MRI-based models. A few years later, Parrella et al.⁴³ also explored models within the same framework, confirming the findings of Buizza et al. to a certain extent. Specifically, the best-performing Parrella et al.’s model also selected shape features, even though texture features were not included. Additionally, the authors introduced dose-averaged linear energy transfer (LETd) maps and physical dose maps-derived radiomics features in their models. These features showed promising results both when fed to the s-SVM model and when integrated into the tumor control probability (TCP) model compared to a clinically based TCP. Despite such encouraging results, it is worth noticing that Buizza et al.’s and Parrella et al.’s studies were conducted relying on a sample gathered in the same institution and under the same protocol. Consequently, further investigations and external validation are needed. Morelli et al.⁴⁴ also investigated biological dose maps and LETd in survival prediction models (s-SVM and r-Cox) to predict local recurrence in sacral chordoma after CIRT. LETd-based radiomics r-Cox was reported as the best model, outperforming the more conventional ones based on dose-volume-histogram (DVH) parameters. Biological dosimetric maps were computed by means of two radiobiological models: the local effect model (LEM v.1) and the micro-dosimetric kinetic model (MKM). Interestingly, only MKM-based dosimetric radiomics led to significant results in the model performances.

Only one of the selected studies was not related to radiomics. Qiu et al.⁴⁵ compared the performance of statistical and ML models in the prediction of high-grade glioma (HGG) relapse after particle therapy with either proton or carbon ion. Using patient, tumor, and treatment variables as input, the Cox model showed better results compared to a random survival forest.

Normal tissue complication prediction

Four research papers, published from 2018 to 2025, explored the application of AI in predicting normal tissue complication prediction (NTCP). The majority of these papers (three out of four) developed their models using clinical, tumor, and treatment data, rather than incorporating radiomics-based features. Additionally, all the studies utilized traditional ML techniques, mostly logistic regression, instead of employing advanced AI models for building NTCP. Although these NTCP models have not undergone external validation, overall, they were characterized by larger sample sizes, ranging from several hundred to over a thousand patients.

In 2018, Zhang et al.⁴⁶ aimed to build a model predicting weight loss in cancer patients treated with particle therapy. The study sample included 365 patients, where no constraints on either the treatment site or particle were applied. Logistic regression and chi-square automatic interaction detector (CHAID) decision tree models were investigated with the following input variables: demographic, nutrition status, tumor-related, treatment-related, and laboratory test results. Basic univariable-based selection methods were implemented to define the multivariable predictors. Both models reported good performances with slightly better accuracy and AUROC for the CHAID decision trees. Li et al.⁴⁷ built a model for predicting high-grade acute radiation dermatitis using 187 head and neck patients treated with CIRT. Although the input predictors were consistent with the classification task and most prior literature (patient-specific and treatment-related factors), the model development presented a systematic structure, with no substantial novelty. Nevertheless, robust procedures were implemented from the feature selection to the model evaluation step. The authors also highlighted that the multivariable model did not outperform the univariable logistic regression models based on RBE-weighted or physical dose to surface parameters. Much later in 2025, Zhang et al.⁴⁸ evaluated the performance of ML approaches in predicting high-grade xerostomia in head and neck cancer patients receiving proton and carbon ion RT. The sample size for model development was remarkable, with 1769 patients, although the heterogeneity and the treatment may hinder the model's specificity to CIRT. The pipeline leading to model development and evaluation was consistent with good practices in ML. After evaluating the most common ML models, linear-SVM reported the highest balanced accuracy, outperforming logistic regression. The authors reported synthetic minority oversampling technique (SMOTE) as a key method to improve the reliability and model performance while accounting for class imbalance. Finally, Meng et al.⁴⁹ developed an NTCP model for predicting high-grade acute oral mucositis in head and neck cancer patients undergoing CIRT. The authors presented the integration of radiomics and dosiomics features in their NTCP model as a major novelty. A conventional logistic regression model was employed, robustly developed and evaluated, achieving a satisfactory AUROC consistent with the literature on the topic. The last three aforementioned studies were evaluated by taking advantage of cross-validation and bootstrap procedures at different stages to ensure a higher level of robustness in the final model.

Discussion

To the best of our knowledge, this is the first scoping review exploring the landscape of AI applications in the context of CIRT. In addition to the two independent human reviewers, the integration of an LLM-based AI reviewer into the screening and data extraction processes represents a novelty in scoping reviews. Although the AI reviewer tended to be slightly liberal in article selection compared to humans (moderate precision), nearly all of the articles selected by humans were also included in the LLM reviewer’s selections (high sensitivity and specificity). In fact, LLM reviewer’s choices were substantially similar to human selections, with no statistically significant difference. While underscoring the immaturity of such tools in conducting independent scoping reviews, our results are encouraging and highlight their potential as a support tool. Furthermore, given the rapid advances in LLMs, we expect higher screening accuracy in the future.

The number of papers meeting the inclusion criteria of our scoping review, in the period between January 2014 and June 2025, was substantially low compared to recent literature reviews on AI applications in proton³ or photon therapy^2,50. Of course, the limited worldwide adoption of CIRT may explain such a discrepancy, leading to fewer AI applications and potentially lower data availability for AI development.

To date, AI research efforts dedicated to CIRT have focused on optimizing laborious CIRT-specific tasks. In fact, DL-based MC dose simulation prediction, denoising, or efficient dose grid sampling were reported to significantly speed up the treatment planning phase, complying with adaptive CIRT requirements. Similarly, the need for dedicated re-planning CT scans during the CIRT course represents additional radiation exposure and computational time. Especially in the framework of adaptive CIRT, such limitations could be impairing; hence, synthetic imaging was identified as a promising tool to aid these processes. sCT from CBCT or MRI offers a promising avenue for low or radiation-free efficient re-planning and for non-invasively monitoring anatomical changes, though clinical translation remains hindered by variations in experimental setup, heterogeneity of anatomical regions, small sample sizes, and limited validation. In this context, model transferability may be a promising research direction to explore.

Indeed, specific AI tools developed in the context of photon⁵¹ or proton⁵² RT could be transferred to CIRT; however, it should be considered that the highest accuracy required in heavy ion therapy demands caution in this process. Dedicated experiments should be designed to validate the existing models and evaluate the need for further tuning, improvement, or implementation of new CIRT-specific AI tools. For example, in the context of CIRT, only one study investigated the feasibility of generating sCT from daily positioning CBCT. Conversely, such an approach was widely explored in proton and photon RT,^53–55 due to a multiplicity of opportunities. Generating sCT from CBCT would lead to a relevant reduction of the facility costs and invasiveness associated with further re-planning CT scans in adaptive CIRT.³⁹ The acceptability of such AI methods should be systematically evaluated with a CIRT-based treatment plan that incorporates dosimetric accuracy constraints, an aspect missing in the aforementioned study.³⁹ On the other hand, an in-silico study by Klassen et al.⁴⁰ reported dosimetric advantages of using CBCT adaptive planning in pancreatic hypo-fractionated CIRT, although the authors did not take advantage of AI-based sCT generation. Moreover, it is also worth noticing that the higher sensitivity to variations along the beam path in CIRT may also result in lower gamma pass rates when compared to photons and protons (~99%), and a consensus for the evaluation of acceptability in CIRT is yet to be established. This further underscores the need for dedicated CIRT experiments on the topic and warrants caution in the direct transferability of such applications to the CIRT domain. From a technical point of view, the AI models considered in CIRT literature were aligned with the most recent literature for RT and PT. No clear trends in the AI approaches over time were identified; in fact, the choice of the AI method seemed to be mostly guided by technical aspects rather than technological availability.

Conversely, we suspect that predictive models may exhibit lower transferability from photon/proton therapy due to CIRT-specific relevance of data modalities (e.g., LETd maps) and heterogeneity in radiobiological models⁵⁰ (LEM vs. MKM, both available in multiple versions) used for dose calculation. While in photons no biological model needs to be applied, and in protons, most of the centers apply a fixed RBE constant (1.1),⁵⁶ in CIRT, different radiobiological models are currently in use without a homogeneous landscape. This underscores a further need for dedicated approaches for CIRT in TCP modeling while emphasizing the difficulty in directly transferring AI models from proton and photon RT. Despite the recent developments in DL, most of the included CIRT studies for prediction tasks (i.e. TCP and NTCP) involved simple ML approaches (such as SVM, s-SVM, or random survival forests) rather than complex AI models (e.g. DL). Such a choice may be justified, given the small sample size, which could generate unstable results when used with data-hungry DL models for prediction tasks. With respect to prediction tasks, it is also worth noticing that 60% of the studies were conducted by the same group, potentially explaining to some extent the methodological homogeneity. Compared to TCP tasks, a trend over time was observed in NTCP modelling. Most of the studies used logistic regression models for NTCP prediction, although the ones published later considered a wider variety of ML models⁴⁸ or included radiomics and dosiomics-derived features in the model. Such a pattern is well aligned with RT literature, where logistic regression models are widely recognized and even validated, making it more difficult to experiment with advanced ML methodologies. Note that, even in the context of proton and photon RT, the application of AI in prediction tasks is currently far from clinical practice; this gap seems wider for CIRT.

Given the scarcity of literature on AI applications in CIRT, this scoping review underscored AI investigation efforts in CIRT, retracing the PT research mainstream. However, a recent literature review on AI in PT³ highlighted additional research areas, including image contouring and image quality improvement, that warrant further attention. As previously discussed, the transferability of most of the proposed approaches is worth evaluating in a CIRT setting and eventually fine-tuned. On the other hand, there are AI-based commercial tools that are currently used in clinical routine,^57,58 mainly for automatic contouring (e.g., for organs of interest), which explains the lack of dedicated studies on the topic. However, the highest CIRT demand for spatial accuracy poses relevant challenges to target segmentation tool development. Most of all, efforts on the evaluation of AI tools to derive mass density and relative stopping power maps from dual energy CT images may be of particular interest in CIRT.⁵⁹ While not extensively explored also in photon and proton RT, AI tools for patient selection are also worth exploring in CIRT, especially considering the fast-growing data digitalization with an increasing widespread use of electronic health records (EHR).

All of the included studies suffered from limited sample size and lacked external validation. The few studies that reported higher samples demonstrated treatment heterogeneity, i.e., combining patients treated either with proton or carbon ion RT. Although for synthetic imaging tasks, with respect to specific evaluation metrics,³⁹ it may not affect the study outcome, we expect a higher impact for prediction tasks. Overall, the highest sample sizes were registered for the NTCP prediction category, while the lowest were for treatment planning and imaging applications. This pattern is consistent with the AI tasks and the available literature. Reduced sample sizes and the lack of external validations lead to low robustness and generalizability of AI models. To mitigate such limitations, we encourage the development of a collaborative network among CIRT centers worldwide to facilitate data sharing, AI models validation, and protocol standardization. However, the intrinsic heterogeneity in CIRT treatment planning approaches, leveraging different radiobiological models, further hinders the collection of large datasets even in a collaborative scenario.

Strengths and limitations

To the best of our knowledge, this is the first scoping review that systematically explored the current landscape of AI in CIRT while critically providing an overview of challenges and future directions. The integration of an LLM-based AI reviewer into the screening and data extraction process presents a novel methodological contribution and could enhance the scale and efficiency of future reviews. Among the limitations of this scoping review is that the research was conducted only on three databases over a 10-year period (2014–2025) and only included articles written in the English language. Nonetheless, we truly believe that the selected articles capture the present AI applications in CIRT and would not be substantially affected by a broader search strategy. Finally, it is worth noting that the literature included in this review reported positive results. In this context, a positive bias may exist; however, reduced data availability and limited widespread use of CIRT to date may partially explain this occurrence.

Conclusions

This scoping review highlights the immaturity of AI developments in the CIRT domain. Besides the limited AI contributions in the field, current studies are further constrained by small sample sizes and a lack of external validation. Despite these limitations, transferring existing AI tools from photon and proton RT to CIRT represents a promising avenue, although caution is warranted in this delicate process. Leveraging AI applications in CIRT holds promise for reducing costs and increasing treatment effectiveness and efficiency, especially in adaptive CIRT scenarios. To support the development of robust and generalizable AI tools in CIRT, the establishment of a collaborative network among CIRT centers is strongly encouraged.

Supplemental Material

sj-docx-1-tmj-10.1177_03008916261438952 – Supplemental material for The landscape of artificial intelligence in carbon ion radiotherapy: A scoping review

Supplemental material, sj-docx-1-tmj-10.1177_03008916261438952 for The landscape of artificial intelligence in carbon ion radiotherapy: A scoping review by Giulia Fontana, Sithin Thulasi Seetha, Sara Lillo, Silvia Molinelli, Alessandro Vai, Anna Maria Camarda, Lorenzo Preda, Amelia Barcellini and Ester Orlandi in Tumori Journal

Footnotes

Author contribution

Conception and design: Ester Orlandi and Amelia Barcellini; Collection and assembly of data: Giulia Fontana and Sithin Thulasi Seetha; Data analysis and interpretation: Giulia Fontana and Sithin Thulasi Seetha; Writing – original draft: Giulia Fontana and Sithin Thulasi Seetha; Writing – review & editing: All authors.

This manuscript has been read and approved by all the authors, and each author believes that the manuscript represents honest work.

Data availability statement

No additional datasets were generated or analyzed in this study. The code supporting the large language model (LLM)-related findings of this study is openly available to support future systematic reviews leveraging fully open-source, locally run LLM for screening and critical appraisal. The source code is available on GitHub: (accessed on 22 January 2026).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Giulia Fontana

Sithin Thulasi Seetha

Sara Lillo

Alessandro Vai

Lorenzo Preda

Amelia Barcellini

Ester Orlandi

Supplemental material

Supplemental material for this article is available online.

References

Guo

Hao

Zhao

, et al. Artificial intelligence in health care: bibliometric analysis. J Med Internet Res 2020; 22: e18228.

Jeong

Goh

Kwak

Challenges and opportunities to integrate artificial intelligence in radiation oncology: a narrative review. Ewha Med J 2024; 47: e49.

Isaksson

Mastroleo

Vincini

, et al. The emerging role of Artificial Intelligence in proton therapy: A review. Crit Rev Oncol/Hematol 2024; 204: 104485.

Ronneberger

Fischer

Brox

. U-Net: Convolutional networks for biomedical image segmentation. In: Navab

Hornegger

Wells

, et al. (eds) Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Cham: Springer International Publishing, pp. 234–241.

Søbstad

Sulen

Pettersen

HES

, et al. Time efficiency, geometric accuracy, and clinical impact of AI-assisted contouring of organs at risk in head and neck cancer radiotherapy. AO 2025; 64: 1194–1201.

Guo

Huang

Yang

, et al. Longitudinal evaluation of workflow optimization in radiotherapy: A 4-year retrospective study. J Applied Clin Med Phys 2025; 26: e70252.

Huet-Dastarac

Michiels

Rivas

, et al. Patient selection for proton therapy using Normal Tissue Complication Probability with deep learning dose prediction for oropharyngeal cancer. Medical Physics 2023; 50: 6201–6214.

Gianoli

De Bernardi

Parodi

‘Under the hood’: artificial intelligence in personalized radiotherapy. BJR Open 2024; 6: tzae017.

Tricco

Lillie

Zarin

, et al. PRISMA Extension for Scoping Reviews (PRISMA-ScR): Checklist and Explanation. Ann Intern Med 2018; 169: 467–473.

10.

Peters

Godfrey

McInerney

Munn

Tricco

Khalil

. Chapter 11: scoping reviews. JBI manual for evidence synthesis. 2020 May 24. Available from https://reviewersmanual.joannabriggs.org/. https://doi.org/10.46658/JBIRM-19-01

11.

Ouzzani

Hammady

Fedorowicz

, et al. Rayyan—a web and mobile app for systematic reviews. Syst Rev 2016; 5: 210.

12.

Guo

Yang

Zhang

Song

Wang

Zhu

Zhang

. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948. 2025 Jan 22.

13.

McHugh

ML.

Interrater reliability: the kappa statistic. Biochem Med (Zagreb) 2012; 22: 276–282.

14.

Waskom

Botvinnik

O’Kane

, et al. mwaskom/seaborn: v0. 8.1 (September 2017). Zenodo. 2017 Sep 3.

15.

Oesper

Merico

Isserlin

, et al. WordCloud: a Cytoscape plugin to create a visual semantic summary of networks. Source Code Biol Med 2011; 6: 7.

16.

Ollama. Ollama homepage. 2025, https://ollama.com (accessed 16 October 2025).

17.

LangChain. LangChain homepage. 2025, https://www.langchain.com (accessed 16 October 2025).

18.

Haddaway

Page

Pritchard

, et al. PRISMA2020 : An R package and Shiny app for producing PRISMA 2020-compliant flow diagrams, with interactivity for optimised digital transparency and Open Synthesis. Campbell Systematic Reviews 2022; 18: e1230.

19.

Yabe

Yamamoto

Oda

, et al. Prediction of dose distribution from luminescence image of water using a deep convolutional neural network for particle therapy. Medical Physics 2020; 47: 3882–3891.

20.

Yabe

Komori

Toshito

, et al. Estimation and correction of produced light from prompt gamma photons on luminescence imaging of water for proton therapy dosimetry. Phys Med Biol 2018; 63: 04NT02.

21.

Yamaguchi

Liu

Huang

, et al. Dose image prediction for range and width verifications from carbon ion-induced secondary electron bremsstrahlung x-rays using deep learning workflow. Medical Physics 2020; 47: 3520–3532.

22.

Zhang

Wang

, et al. Deep learning-based fast denoising of Monte Carlo dose calculation in carbon ion radiotherapy. Medical Physics 2023; 50: 7314–7323.

23.

Muraro

Battistoni

Kraan

AC.

Challenges in Monte Carlo simulations as clinical and research tool in particle therapy: a review. Front Phys 2020; 8: 567800.

24.

Zhu

Park

Isola

, et al. Unpaired Image-to-Image Translation Using Cycle-Consistent Adversarial Networks, in 2017 IEEE International Conference on Computer Vision (ICCV), Venice, Italy, 2017, pp. 2242–2251, doi: 10.1109/ICCV.2017.244.

25.

Çiçek

Abdulkadir

Lienkamp

, et al. (2016). 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation. In: Ourselin

Joskowicz

Sabuncu

Unal

Wells

, (eds) Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016. MICCAI 2016. Lecture Notes in Computer Science, vol 9901. Springer, Cham. https://doi.org/10.1007/978-3-319-46723-8_49

26.

Kazerouni

Dooly

Toal

Ghost-UNet: An asymmetric encoder-decoder architecture for semantic segmentation from scratch. IEEE Access 2021; 9: 97457–97465.

27.

Javaid

Souris

Dasnoy

, et al. Mitigating inherent noise in Monte Carlo dose distributions using dilated U-Net. Medical Physics 2019; 46: 5790–5798.

28.

Zhang

Chen

Dai

, et al. A plan verification platform for online adaptive proton therapy using deep learning-based Monte–Carlo denoising. Physica Medica 2022; 103: 18–25.

29.

Wang

, et al. Deep learning-based Monte Carlo dose prediction for heavy-ion online adaptive radiotherapy and fast quality assurance: A feasibility study. Medical Physics 2025; 52: 2570–2580.

30.

Quarz

Volz

Antink

, et al. Deep learning-based voxel sampling for particle therapy treatment planning. Phys Med Biol 2024; 69: 155014.

31.

Martin

Bortfeld

Castañon

DA.

Accelerating IMRT optimization by voxel sampling. Phys Med Biol 2007; 52: 7211–7228.

32.

Wang

Zuluaga

, et al. DeepIGeoS: A deep interactive geodesic framework for medical image segmentation. IEEE Trans Pattern Anal Mach Intell 2019; 41: 1559–1572.

33.

Zhang

, et al. DR-only Carbon-ion radiotherapy treatment planning via deep learning. Physica Medica 2022; 100: 120–128.

34.

Knäusl

Kuess

Stock

, et al. Possibilities and challenges when using synthetic computed tomography in an adaptive carbon-ion treatment workflow. Zeitschrift für Medizinische Physik 2023; 33: 146–154.

35.

Parrella

Vai

Nakas

, et al. Synthetic CT in carbon ion radiotherapy of the abdominal site. Bioengineering 2023; 10: 250.

36.

Nakas

Hladchuk

Parrella

, et al. Deep-learning synthetized 4DCT from 4DMRI of the abdominal site in carbon-ion radiotherapy. Physica Medica 2025; 133: 104963.

37.

Isola

Zhu

, et al. Image-to-Image Translation with Conditional Adversarial Networks, in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, HI, USA, 2017, pp. 5967–5976, doi: 10.1109/CVPR.2017.632.

38.

Florkow

Guerreiro

Zijlstra

, et al. Deep learning-enabled MRI-only photon and proton therapy treatment planning for paediatric abdominal tumours. Radiotherapy and Oncology 2020; 153: 220–227.

39.

Pepa

Taleghani

Sellaro

, et al. Unsupervised deep learning for synthetic CT generation from CBCT images for proton and carbon ion therapy for paediatric patients. Sensors 2024; 24: 7460.

40.

Klassen

Hirashima

Iramina

, et al. Dosimetric performance of cone beam CT-guided adaptive carbon-ion radiotherapy with daily replanning for pancreatic cancer. Physica Medica 2025; 134: 104991.

41.

Jiao

Zhang

, et al. Imaging-based individualized response prediction of carbon ion radiotherapy for prostate cancer patients. CMAR 2019; 11: 9121–9131.

42.

Buizza

Paganelli

D’Ippolito

, et al. Radiomics and dosiomics for predicting local control after carbon-ion radiotherapy in skull-base chordoma. Cancers 2021; 13: 339.

43.

Parrella

Annunziata

Morelli

, et al. A dosiomics approach to treatment outcome modeling in carbon ion radiotherapy for skull base chordomas. Physica Medica 2024; 124: 103421.

44.

Morelli

Parrella

Molinelli

, et al. A dosiomics analysis based on linear energy transfer and biological dose maps to predict local recurrence in sacral chordomas after carbon-ion radiotherapy. Cancers 2022; 15: 33.

45.

Qiu

Gao

Yang

, et al. A comparison study of machine learning (Random Survival Forest) and classic statistic (Cox Proportional Hazards) for predicting progression in high-grade glioma after proton and carbon ion radiotherapy. Front Oncol 2020; 10: 551420.

46.

Zhang

Zhu

Zhang

, et al. Prediction model of critical weight loss in cancer patients during particle therapy. Japan J Clin Oncol 2018; 48: 75–81.

47.

Sakai

Tsunoda

, et al. Normal tissue complication probability model for acute radiation dermatitis in patients with head and neck cancer treated with carbon ion radiation therapy. Int J Radiat Oncol Biol Physic 2022; 113: 675–684.

48.

Zhang

Wang

, et al. Evaluation of machine learning models for predicting xerostomia in adults with head and neck cancer during proton and heavy ion radiotherapy. Radiother Oncol 2025; 204: 110712.

49.

Meng

Sakai

, et al. Normal tissue complication probability model for acute oral mucositis in patients with head and neck cancer undergoing carbon ion radiation therapy based on dosimetry, radiomics, and dosiomics. Radiother Oncol 2025; 204: 110709.

50.

Morelli

Banini

Greto

, et al. Integrating Radiomics and Artificial Intelligence (AI) in Stereotactic Body Radiotherapy (SBRT)/Stereotactic Radiosurgery (SRS): Predictive Tools for Tailored Cancer Care. Cancers 2025; 17: 2906.

51.

Duke

Papanikolaou

Artificial intelligence in radiation therapy: from imaging to delivery—a comprehensive review. BJR|Artificial Intelligence 2025; 2: ubaf012.

52.

Wildman

Wynne

Momin

, et al. Recent advances in applying machine learning to proton radiotherapy. Biomed Phys Eng Express 2025; 11: 042005.

53.

Fischer

Bensberg

, et al. CBCT-based online adaptive radiotherapy of the prostate bed: first clinical experience and comparison to nonadaptive conventional IGRT. Strahlenther Onkol 2025; 201: 767–778.

54.

Rossi

Cerveri

Comparison of supervised and unsupervised approaches for the generation of synthetic CT from cone-beam CT. Diagnostics 2021; 11: 1435.

55.

Vestergaard

Muren

Elstrøm

, et al. Daily proton dose re-calculation on deep-learning corrected cone-beam computed tomography scans. Radiother Oncol 2025; 209: 110953.

56.

Heuchel

Hahn

Pawelke

, et al. Clinical use and future requirements of relative biological effectiveness: Survey among all European proton therapy centres. Radiother Oncol 2022; 172: 134–139.

57.

Doolan

Charalambous

Roussakis

, et al. A clinical evaluation of the performance of five commercial artificial intelligence contouring systems for radiotherapy. Front Oncol 2023; 13: 1213068.

58.

Rusanov

Ebert

Sabet

, et al. Guidance on selecting and evaluating AI auto-segmentation systems in clinical radiotherapy: insights from a six-vendor analysis. Phys Eng Sci Med 2025; 48: 301–316.

59.

Chang

C-W

Gao

Wang

, et al. Dual-energy CT based mass density and relative stopping power estimation for proton therapy using physics-informed deep learning. Phys Med Biol 2022; 67: 115010.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.03 MB