Virtual 4DCT generated from 4DMRI in gated particle therapy: phantom validation and application to lung cancer patients.

Objective. Respiration negatively affects the outcome of a radiation therapy treatment, with potentially severe effects especially in particle therapy (PT). If compensation strategies are not applied, accuracy cannot be achieved. To support the clinical practice based on 4D computed tomography (CT), 4D magnetic resonance imaging (MRI) acquisitions can be exploited. The purpose of this study was to validate a method for virtual 4DCT generation from 4DMRI data for lung cancers on a porcine lung phantom, and to apply it to lung cancer patients in PT.Approach. Deformable image registration was used to register each respiratory phase of the 4DMRI to a reference phase. Then, a static 3DCT was registered to this reference MR image set, and the virtual 4DCT was generated by warping the registered CT according to previously obtained deformation fields. The method was validated on a physical phantom for which a ground truth 4DCT was available and tested on lung tumor patients, treated with gated PT at end-exhale, by comparing the virtual 4DCT with a re-evaluation 4DCT. The geometric and dosimetric evaluation was performed for both proton and carbon ion treatment plans.Main results. The phantom validation exhibited a geometrical accuracy within the maximum resolution of the MRI and mean dose deviations, with respect to the prescription dose, up to 3.2% for targetD95%, with a mean gamma pass rate of 98%. For patients, the virtual and re-evaluation 4DCTs showed good correspondence, with errors on targetD95%up to 2% within the gating window. For one patient, dose variations up to 10% at end-exhale were observed due to relevant inter-fraction anatomo-pathological changes that occurred between the planning and re-evaluation CTs.Significance. Results obtained on phantom data showed that the virtual 4DCT method was accurate, allowing its application on patient data for testing within a clinical scenario.

Integration of spatial distortion effects in a 4D computational phantom for simulation studies in extra-cranial MRI-guided radiation therapy: Initial results.

PURPOSE: Spatial distortions in magnetic resonance imaging (MRI) are mainly caused by inhomogeneities of the static magnetic field, nonlinearities in the applied gradients, and tissue-specific magnetic susceptibility variations. These factors may significantly alter the geometrical accuracy of the reconstructed MR image, thus questioning the reliability of MRI for guidance in image-guided radiation therapy. In this work, we quantified MRI spatial distortions and created a quantitative model where different sources of distortions can be separated. The generated model was then integrated into a four-dimensional (4D) computational phantom for simulation studies in MRI-guided radiation therapy at extra-cranial sites. METHODS: A geometrical spatial distortion phantom was designed in four modules embedding laser-cut PMMA grids, providing 3520 landmarks in a field of view of (345 × 260 × 480) mm3 . The construction accuracy of the phantom was verified experimentally. Two fast MRI sequences for extra-cranial imaging at 1.5 T were investigated, considering axial slices acquired with online distortion correction, in order to mimic practical use in MRI-guided radiotherapy. Distortions were separated into their sources by acquisition of images with gradient polarity reversal and dedicated susceptibility calculations. Such a separation yielded a quantitative spatial distortion model to be used for MR imaging simulations. Finally, the obtained spatial distortion model was embedded into an anthropomorphic 4D computational phantom, providing registered virtual CT/MR images where spatial distortions in MRI acquisition can be simulated. RESULTS: The manufacturing accuracy of the geometrical distortion phantom was quantified to be within 0.2 mm in the grid planes and 0.5 mm in depth, including thickness variations and bending effects of individual grids. Residual spatial distortions after MRI distortion correction were strongly influenced by the applied correction mode, with larger effects in the trans-axial direction. In the axial plane, gradient nonlinearities caused the main distortions, with values up to 3 mm in a 1.5 T magnet, whereas static field and susceptibility effects were below 1 mm. The integration in the 4D anthropomorphic computational phantom highlighted that deformations can be severe in the region of the thoracic diaphragm, especially when using axial imaging with 2D distortion correction. Adaptation of the phantom based on patient-specific measurements was also verified, aiming at increased realism in the simulation. CONCLUSIONS: The implemented framework provides an integrated approach for MRI spatial distortion modeling, where different sources of distortion can be quantified in time-dependent geometries. The computational phantom represents a valuable platform to study motion management strategies in extra-cranial MRI-guided radiotherapy, where the effects of spatial distortions can be modeled on synthetic images in a virtual environment.

Improving the modelling of susceptibility-induced spatial distortions in MRI-guided extra-cranial radiotherapy.

Magnetic-resonance linear-accelerator (MR-LINAC) systems integrating in-room magnetic-resonance-imaging (MRI) guidance are a currently emerging technology. Such systems address the need to provide frequent imaging at optimal soft-tissue contrast for treatment guidance. However, the use of MRI-guidance in radiotherapy should address imaging-related spatial distortions, which may hinder accurate geometrical characterization of the treatment site. Since spatial encoding relies on well-defined magnetic fields, accurate modeling of the magnetic field alterations due to [Formula: see text]-inhomogeneities, gradient nonlinearities, and susceptibilities is needed. In this work, the modeling of susceptibility induced distortions is considered. Dedicated susceptibility measurements are reported, aiming at extending the characterization of different tissues for MRI-guided extra-cranial radiotherapy applications. A digital 4D anthropomorphic phantom, providing time-resolved anatomical changes due to breathing, is exploited as reference anatomy to quantify spatial distortions due to variations in tissue susceptibility. Sub-millimeter values can be attributed to susceptibility-induced distortions, with maximum values up to 2.3 mm at a gradient strength of 5 mT m-1. Improvements in susceptibility simulation for extra-cranial sites are shown when including specifically the contributions from lung, liver and muscular tissues.

Time-resolved volumetric MRI in MRI-guided radiotherapy: an in silico comparative analysis.

MRI-treatment units enable 2D cine-MRI centred in the tumour for motion detection in radiotherapy, but they lack 3D information due to spatio-temporal limits. To derive time-resolved 3D information, different approaches have been proposed in the literature, but a rigorous comparison among these strategies has not yet been performed. The goal of this study is to quantitatively investigate five published strategies that derive time-resolved volumetric MRI in MRI-guided radiotherapy: Propagation, out-of-plane motion compensation, Fayad model, ROI-based model and Stemkens model. Comparisons were performed using an MRI digital phantom generated with six different patient-derived motion signals and tumour-shapes. An average 4D cycle was generated as well as 2D cine-MRI data with corresponding 3D in-room ground truth. Quantitative analysis was performed by comparing the estimated 3D volume to the ground truth available for each 2D cine-MRI sample. A grouped patient statistical analysis was performed to evaluate the performance of the selected methods, in case of tumour tracking or motion estimation of the whole anatomy. Analyses were also performed based on patient characteristics. Quantitative ranking of the investigated methods highlighted that Propagation and ROI-based model strategies achieved an overall median tumour centre of mass 3D distance from the ground truth of 1.1 mm and 1.3 mm, respectively, and a diaphragm distance below 1.6 mm. Higher errors and variabilities were instead obtained for other methods, which lack the ability to compensate for in-room variations and to account for regional changes. These results were especially evident when further analysing patient characteristics, where errors above 2 mm/5 mm in tumour/diaphragm were found for more irregular breathing patterns in case of out-of-plane motion compensation, Fayad and Stemkens models. These findings suggest the potential of the proposed in silico framework to develop and compare strategies to estimate time-resolved 3DMRI in MRI-guided radiotherapy.

MRI-guidance for motion management in external beam radiotherapy: current status and future challenges.

High precision conformal radiotherapy requires sophisticated imaging techniques to aid in target localisation for planning and treatment, particularly when organ motion due to respiration is involved. X-ray based imaging is a well-established standard for radiotherapy treatments. Over the last few years, the ability of magnetic resonance imaging (MRI) to provide radiation-free images with high-resolution and superb soft tissue contrast has highlighted the potential of this imaging modality for radiotherapy treatment planning and motion management. In addition, these advantageous properties motivated several recent developments towards combined MRI radiation therapy treatment units, enabling in-room MRI-guidance and treatment adaptation. The aim of this review is to provide an overview of the state-of-the-art in MRI-based image guidance for organ motion management in external beam radiotherapy. Methodological aspects of MRI for organ motion management are reviewed and their application in treatment planning, in-room guidance and adaptive radiotherapy described. Finally, a roadmap for an optimal use of MRI-guidance is highlighted and future challenges are discussed.

4D ML reconstruction as a tool for volumetric PET-based treatment verification in ion beam radiotherapy.

PURPOSE: An innovative strategy to improve the sensitivity of positron emission tomography (PET)-based treatment verification in ion beam radiotherapy is proposed. METHODS: Low counting statistics PET images acquired during or shortly after the treatment (Measured PET) and a Monte Carlo estimate of the same PET images derived from the treatment plan (Expected PET) are considered as two frames of a 4D dataset. A 4D maximum likelihood reconstruction strategy was adapted to iteratively estimate the annihilation events distribution in a reference frame and the deformation motion fields that map it in the Expected PET and Measured PET frames. The outputs generated by the proposed strategy are as follows: (1) an estimate of the Measured PET with an image quality comparable to the Expected PET and (2) an estimate of the motion field mapping Expected PET to Measured PET. The details of the algorithm are presented and the strategy is preliminarily tested on analytically simulated datasets. RESULTS: The algorithm demonstrates (1) robustness against noise, even in the worst conditions where 1.5 × 10(4) true coincidences and a random fraction of 73% are simulated; (2) a proper sensitivity to different kind and grade of mismatches ranging between 1 and 10 mm; (3) robustness against bias due to incorrect washout modeling in the Monte Carlo simulation up to 1/3 of the original signal amplitude; and (4) an ability to describe the mismatch even in presence of complex annihilation distributions such as those induced by two perpendicular superimposed ion fields. CONCLUSIONS: The promising results obtained in this work suggest the applicability of the method as a quantification tool for PET-based treatment verification in ion beam radiotherapy. An extensive assessment of the proposed strategy on real treatment verification data is planned.

Motion prediction in MRI-guided radiotherapy based on interleaved orthogonal cine-MRI.

In-room cine-MRI guidance can provide non-invasive target localization during radiotherapy treatment. However, in order to cope with finite imaging frequency and system latencies between target localization and dose delivery, tumour motion prediction is required. This work proposes a framework for motion prediction dedicated to cine-MRI guidance, aiming at quantifying the geometric uncertainties introduced by this process for both tumour tracking and beam gating. The tumour position, identified through scale invariant features detected in cine-MRI slices, is estimated at high-frequency (25 Hz) using three independent predictors, one for each anatomical coordinate. Linear extrapolation, auto-regressive and support vector machine algorithms are compared against systems that use no prediction or surrogate-based motion estimation. Geometric uncertainties are reported as a function of image acquisition period and system latency. Average results show that the tracking error RMS can be decreased down to a [0.2; 1.2] mm range, for acquisition periods between 250 and 750 ms and system latencies between 50 and 300 ms. Except for the linear extrapolator, tracking and gating prediction errors were, on average, lower than those measured for surrogate-based motion estimation. This finding suggests that cine-MRI guidance, combined with appropriate prediction algorithms, could relevantly decrease geometric uncertainties in motion compensated treatments.

Adaptive Mathematical Model of Tumor Response to Radiotherapy Based on CBCT Data.

Mathematical modeling of tumor response to radiotherapy has the potential of enhancing the quality of the treatment plan, which can be even tailored on an individual basis. Lack of extensive in vivo validation has prevented, however, reliable clinical translation of modeling outcomes. Image-guided radiotherapy is a consolidated treatment modality based on computed tomographic (CT) imaging for tumor delineation and volumetric cone beam CT data for periodic checks during treatment. In this study, a macroscopic model of tumor growth and radiation response is proposed, being able to adapt along the treatment course as volumetric tumor data become available. Model parameter learning was based on cone beam CT images in 13 uterine cervical cancer patients, subdivided into three groups (G1, G2, G3) according to tumor type and treatment. Three group-specific parameter sets (PS1, PS2, and PS3) on one general parameter set (PSa) were applied. The corresponding average model fitting errors were 14%, 18%, 13%, and 21%, respectively. The model adaptation testing was performed using volume data of three patients, other than the ones involved in the parameter learning. The extrapolation performance of the general model was improved, while comparable prediction errors were found for the group-specific approach. This suggests that an online parameter tuning can overcome the limitations of a suboptimal patient stratification, which appeared otherwise a critical issue.

Challenges and opportunities in image guided particle therapy.

The application of biomedical imaging and image processing to radiation therapy with accelerated particles has unique challenges. The potential of particle therapy to precisely tailor the dose distribution around the target volume needs to account for the intrinsic sensitivity to uncertainties in dose deposition. These peculiar features motivate the use of image guided methods to consistently verify the accuracy in dose delivery. Dedicated imaging and image processing methods are required, from treatment planning to treatment verification phases, in order to reduce the effects of uncertainties. The scenario is also complicated by the lack of standardized layouts of treatment bunkers, which implies the relatively increased use of custom solutions. Conversely, imaging can be applied to verify the actual delivered dose, representing a valuable opportunity to validate specific protocols and visualize the efficacy of the intended treatment. In this contribution, challenges and opportunities in image guided particle therapy are overviewed, with a clear focus on research perspectives in biomedical imaging and image processing.

Image guided particle therapy in CNAO room 2: implementation and clinical validation.

In this contribution we describe the implementation of a novel solution for image guided particle therapy, designed to ensure the maximal accuracy in patient setup. The presented system is installed in the central treatment room at Centro Nazionale di Adroterapia Oncologica (CNAO, Italy), featuring two fixed beam lines (horizontal and vertical) for proton and carbon ion therapy. Treatment geometry verification is based on robotic in-room imaging acquisitions, allowing for 2D/3D registration from double planar kV-images or 3D/3D alignment from cone beam image reconstruction. The calculated six degrees-of-freedom correction vector is transferred to the robotic patient positioning system, thus yielding automated setup error compensation. Sub-millimetre scale residual errors were measured in absolute positioning of rigid phantoms, in agreement with optical- and laser-based assessment. Sub-millimetre and sub-degree positioning accuracy was achieved when simulating setup errors with anthropomorphic head, thorax and pelvis phantoms. The in-house design and development allowed a high level of system customization, capable of replicating the clinical performance of commercially available products, as reported with preliminary clinical results in 10 patients.

Commissioning and quality assurance of an integrated system for patient positioning and setup verification in particle therapy.

In an increasing number of clinical indications, radiotherapy with accelerated particles shows relevant advantages when compared with high energy X-ray irradiation. However, due to the finite range of ions, particle therapy can be severely compromised by setup errors and geometric uncertainties. The purpose of this work is to describe the commissioning and the design of the quality assurance procedures for patient positioning and setup verification systems at the Italian National Center for Oncological Hadrontherapy (CNAO). The accuracy of systems installed in CNAO and devoted to patient positioning and setup verification have been assessed using a laser tracking device. The accuracy in calibration and image based setup verification relying on in room X-ray imaging system was also quantified. Quality assurance tests to check the integration among all patient setup systems were designed, and records of daily QA tests since the start of clinical operation (2011) are presented. The overall accuracy of the patient positioning system and the patient verification system motion was proved to be below 0.5 mm under all the examined conditions, with median values below the 0.3 mm threshold. Image based registration in phantom studies exhibited sub-millimetric accuracy in setup verification at both cranial and extra-cranial sites. The calibration residuals of the OTS were found consistent with the expectations, with peak values below 0.3 mm. Quality assurance tests, daily performed before clinical operation, confirm adequate integration and sub-millimetric setup accuracy. Robotic patient positioning was successfully integrated with optical tracking and stereoscopic X-ray verification for patient setup in particle therapy. Sub-millimetric setup accuracy was achieved and consistently verified in daily clinical operation.

Regularization in deformable registration of biomedical images based on divergence and curl operators.

BACKGROUND: Similarity measures in medical images do not uniquely determine the correspondence between two voxels in deformable image registration. Uncertainties in the final computed deformation exist, questioning the actual physiological consistency of the deformation between the two images. OBJECTIVES: We developed a deformable image registration method that regularizes the deformation field in order to model a deformation with physiological properties, relying on vector calculus based operators as a regularization function. METHOD: We implemented a 3D multi-resolution parametric deformable image registration, containing divergence and curl of the deformation field as regularization terms. Exploiting a BSpline model, we fit the transformation to optimize histogram-based mutual information similarity measure. In order to account for compression/expansion, we extract sink/source/circulation components as irregularities in the warped image and compensate them. The registration performance was evaluated using Jacobian determinant of the deformation field, inverse-consistency, landmark errors and residual image difference along with displacement field errors. Finally, we compare our results to a robust combination of second derivative regularization, as well as to non-regularized methods. RESULTS: The implementation was tested on synthetic phantoms and clinical data, leading to increased image similarity and reduced inverse-consistency errors. The statistical analysis on clinical cases showed that regularized methods are able to achieve better image similarity than non regularized methods. Also, divergence/curl regularization improves anatomical landmark errors compared to second derivative regularization. CONCLUSION: The implemented divergence/curl regularization was successfully tested, leading to promising results in comparison with competitive regularization methods. Future work is required to establish parameter tuning and reduce the computational cost.

Commissioning of an integrated platform for time-resolved treatment delivery in scanned ion beam therapy by means of optical motion monitoring.

The integrated use of optical technologies for patient monitoring is addressed in the framework of time-resolved treatment delivery for scanned ion beam therapy. A software application has been designed to provide the therapy control system (TCS) with a continuous geometrical feedback by processing the external surrogates tridimensional data, detected in real-time via optical tracking. Conventional procedures for phase-based respiratory phase detection were implemented, as well as the interface to patient specific correlation models, in order to estimate internal tumor motion from surface markers. In this paper, particular attention is dedicated to the quantification of time delays resulting from system integration and its compensation by means of polynomial interpolation in the time domain. Dedicated tests to assess the separate delay contributions due to optical signal processing, digital data transfer to the TCS and passive beam energy modulation actuation have been performed. We report the system technological commissioning activities reporting dose distribution errors in a phantom study, where the treatment of a lung lesion was simulated, with both lateral and range beam position compensation. The zero-delay systems integration with a specific active scanning delivery machine was achieved by tuning the amount of time prediction applied to lateral (14.61 ± 0.98 ms) and depth (34.1 ± 6.29 ms) beam position correction signals, featuring sub-millimeter accuracy in forward estimation. Direct optical target observation and motion phase (MPh) based tumor motion discretization strategies were tested, resulting in 20.3(2.3)% and 21.2(9.3)% median (IQR) percentual relative dose difference with respect to static irradiation, respectively. Results confirm the technical feasibility of the implemented strategy towards 4D treatment delivery, with negligible percentual dose deviations with respect to static irradiation.

Tumor tracking based on correlation models in scanned ion beam therapy: an experimental study.

Accurate dose delivery to extra-cranial lesions requires tumor motion compensation. An effective compensation can be achieved by real-time tracking of the target position, either measured in fluoroscopy or estimated through correlation models as a function of external surrogate motion. In this work, we integrated two internal/external correlation models (a state space model and an artificial neural network-based model) into a custom infra-red optical tracking system (OTS). Dedicated experiments were designed and conducted at GSI (Helmholtzzentrum für Schwerionenforschung). A robotic breathing phantom was used to reproduce regular and irregular internal target motion as well as external thorax motion. The position of a set of markers placed on the phantom thorax was measured with the OTS and used by the correlation models to infer the internal target position in real-time. Finally, the estimated target position was provided as input for the dynamic steering of a carbon ion beam. Geometric results showed that the correlation models transversal (2D) targeting error was always lower than 1.3 mm (root mean square). A significant decrease of the dosimetric error with respect to the uncompensated irradiation was achieved in four out of six experiments, demonstrating that phase shifts are the most critical irregularity for external/internal correlation models.

Validation of automatic contour propagation for 4D treatment planning using multiple metrics.

The aim of this work is to provide insights into multiple metrics clinical validation of deformable image registration and contour propagation methods in 4D lung radiotherapy planning. The following indices were analyzed and compared: Volume Difference (VD), Dice Similarity Coefficient (DSC), Positive Predictive Value (PPV) and Surface Distances (SD). The analysis was performed on three patient datasets, using as reference a ground-truth volume generated by means of Simultaneous Truth And Performance Level Estimation (STAPLE) algorithm from the outlines of five experts. Significant discrepancies in the quality assessment provided by the different metrics in all the examined cases were found. Metrics sensitivity was more evident in presence of image artifacts and particularly for tubular anatomical structures, such as esophagus or spinal cord. Volume Differences did not account for position and DSC exhibited criticalities due to its intrinsic symmetry (i.e. over- and under-estimation of the reference contours cannot be discriminated) and dependency on the total volume of the structure. PPV analysis showed more robust performance, as each voxel concurs to the classification of the propagation, but was not able to detect inclusion of propagated and ground-truth volumes. Mesh distances could interpret the actual shape of the structures, but might report higher mismatches in case of large local differences in the contour surfaces. According to our study, the combination of VD and SD for the validation of contour propagation algorithms in 4D could provide the necessary failure detection accuracy.

The role of regularization in deformable image registration for head and neck adaptive radiotherapy.

Deformable image registration provides a robust mathematical framework to quantify morphological changes that occur along the course of external beam radiotherapy treatments. As clinical reliability of deformable image registration is not always guaranteed, algorithm regularization is commonly introduced to prevent sharp discontinuities in the quantified deformation and achieve anatomically consistent results. In this work we analyzed the influence of regularization on two different registration methods, i.e. B-Splines and Log Domain Diffeomorphic Demons, implemented in an open-source platform. We retrospectively analyzed the simulation computed tomography (CTsim) and the corresponding re-planning computed tomography (CTrepl) scans in 30 head and neck cancer patients. First, we investigated the influence of regularization levels on hounsfield units (HU) information in 10 test patients for each considered method. Then, we compared the registration results of the open-source implementation at selected best performing regularization levels with a clinical commercial software on the remaining 20 patients in terms of mean volume overlap, surface and center of mass distances between manual outlines and propagated structures. The regularized B-Splines method was not statistically different from the commercial software. The tuning of the regularization parameters allowed open-source algorithms to achieve better results in deformable image registration for head and neck patients, with the additional benefit of a framework where regularization can be tuned on a patient specific basis.

Robustness of external/internal correlation models for real-time tumor tracking to breathing motion variations.

In radiotherapy, organ motion mitigation by means of dynamic tumor tracking requires continuous information about the internal tumor position, which can be estimated relying on external/internal correlation models as a function of external surface surrogates. In this work, we propose a validation of a time-independent artificial neural networks-based tumor tracking method in the presence of changes in the breathing pattern, evaluating the performance on two datasets. First, simulated breathing motion traces were specifically generated to include gradually increasing respiratory irregularities. Then, seven publically available human liver motion traces were analyzed for the assessment of tracking accuracy, whose sensitivity with respect to the structural parameters of the model was also investigated. Results on simulated data showed that the proposed method was not affected by hysteretic target trajectories and it was able to cope with different respiratory irregularities, such as baseline drift and internal/external phase shift. The analysis of the liver motion traces reported an average RMS error equal to 1.10 mm, with five out of seven cases below 1 mm. In conclusion, this validation study proved that the proposed method is able to deal with respiratory irregularities both in controlled and real conditions.

Automated fiducial localization in CT images based on surface processing and geometrical prior knowledge for radiotherapy applications.

We propose a novel method for radio-opaque external marker localization in CT scans for infrared (IR) patient set-up in radiotherapy. Efforts were focused on the quantification of uncertainties in marker localization in the CT dataset as a function of algorithm performance. We implemented a 3-D approach to fiducial localization based on surface extraction and marker recognition according to geometrical prior knowledge. The algorithm parameters were optimized on a clinical CT dataset coming from 35 cranial and extra-cranial patients; the localization accuracy was benchmarked at variable image resolution versus laser tracker measurements. The applicability of conventional IR optical tracking systems for localizing external surrogates in daily patient set-up procedures was also investigated in 121 proton therapy treatment sessions. Our study shows that the implemented algorithm features surrogates localization with uncertainties lower than 0.3 mm and with a true positive rate of 90.1%, being this latter mainly influenced by fiducial homogeneity in the CT images. The reported clinical validation in proton therapy confirmed the submillimetric accuracy and the expected algorithm sensitivity. Geometrical prior knowledge allows judging the reliability of the extracted fiducial coordinates, ensuring the highest accuracy in patient set-up.

Scale Invariant Feature Transform as feature tracking method in 4D imaging: a feasibility study.

We propose the use of Scale Invariant Feature Transform (SIFT) as a method able to extract stable landmarks from 4D images and to quantify internal motion. We present a preliminary validation of the SIFT method relying on expert user identification of landmarks and then apply it to 4D lung CT and liver MRI data. Results demonstrate SIFT capabilities as an operator-independent feature tracking method.

Development and validation of a prototypal neural networks-based tumor tracking method.

In radiotherapy, intra-fractional organ motion introduces uncertainties in target localization, leading to unacceptable inaccuracy in dose delivery. Especially in highly selective treatments, such as those delivered with particles beams instead of photons, organ motion may results in severe side effects and/or limited tumor control. Tumor tracking is a motion mitigation strategy that allows an almost continuous dose delivery while the beam is dynamically steered to match the position of the moving target in real-time. Currently, tumor tracking is applied clinically only in the CyberKnife system for photon radiotherapy, whereas neither clinical solutions nor dedicated methodologies are available for particle therapy. Consequently, the aim of the proposed study is to develop a neural networks-based prototypal tracking algorithm intended for particle therapy. We developed a method that exploits three independent neural networks to estimate the internal target position as a function of external surrogate signals. This method was tested on data relative to 20 patients treated with CyberKnife, whose performance was used as benchmark. Results show that the developed algorithm allows targeting error reduction with respect to the CyberKnife system, thus proving the potential value of artificial neural networks for the implementation of tumor tracking methodologies.