Dosimetric tool for assessing lymphocyte pool irradiation

Abstract

The present invention relates to a method for determining the irradiation dose that has been or will be received by a lymphocyte L during a radiotherapy treatment performed on a target organ of a patient, said method comprising the steps of: a) collecting the following data associated to said radiotherapy treatment: - the radiation treatment planning applied to or scheduled for said patient, - the radiation reference doses applied to or scheduled for said patient, and - the number, location and volume of the lymphoid organs that have been or will be irradiated (if any), b) simulating the lymphocyte path between the several organs of the body, c) determining the irradiation dose received by said lymphocyte L during said radiotherapy treatment would it be circulating in peripheral blood, or present in or circulating between lymphoid organs, d) optionally, determining the irradiation dose received by at least one out-of-field lymphoid organ during said radiotherapy treatment, e) combining these data in order to determine the irradiation dose that has been or will be received by the lymphocytes of said patient during said radiotherapy treatment.

Description

[0001] DOSIMETRIC TOOL FOR ASSESSING LYMPHOCYTE POOL IRRADIATION

[0002] DESCRIPTION OF THE PRIOR ART

[0003] Lymphocytes play a central role in the response to radio-chemotherapy and immunoradiotherapy [1], However, one of the iatrogenic effects commonly observed in radiotherapy (RT) is the appearance of severe radiation-induced lymphopenia (sRIL), which is defined as a significant decrease in the absolute lymphocyte count (ALC) under 500 lymphocytes per mm3of blood [2], As an example, approximately 30 to 40% of sRIL are observed during or after the end of RT treatment in patients with gliomas [3], [4], sRIL can persist for several months after the end of RT and represents an independent risk factor for reduced overall survival (OS) for both radio-chemotherapy and immuno-radiotherapy treatments ([5]), with for instance an estimated hazard ratio of 1.86 for patients with glioblastoma [6], 2.94 for patients with head and neck cancer [7] and 1.59 for patients with lung cancer [8], Several independent variables have been identified as bad prognostic of sRIL such as a large volume of irradiation, a normo-fractionated treatment (compared with hypo-fractionated), photon irradiation (compared with protons) or the use of concomitant chemotherapy [5]). Recently, dosimetric constraints for immunological organs have been compiled [9], with the goal to move to lymphocytesparing RT strategies to improve treatment outcomes by enhancing the anti-tumor immune response

[0010] .

[0004] In this context, it is essential to develop tools to assess the dose received by lymphocytes during radiotherapy treatment. Several research groups have been interested in developing models to specifically assess doses received by blood lymphocytes (

[0013] ;

[0021] ;

[0016] ;

[0014] ;

[0011] ;

[0026] ; ). This could be particularly relevant for brain tumors that contain few lymphoid structures and progenitor cells

[0017] , and where many of the irradiated lymphocytes in the RT field are found in circulating blood.

[0005] These teams previously intended to simulate radiation doses received by circulating lymphocytes, without considering recirculation or Out-Of-Field (OOF) doses

[0012] ,

[0013] , However, accurate estimation of the radiation dose delivered to lymphocytes is a complex task considering lymphocytes' dynamic within the body. Lymphocytes can traffic along lymph vessels, in peripheral blood and also be transiently retained in lymphoid organs by specific biological mechanisms

[0018] ,

[0019] , thus transiently exiting the blood circulation and reintegrating it later, implying that the whole bloodcirculating lymphocyte pool can be renewed up to 11 times a day. Thus, peripheral lymphocytes present in the blood at the time of irradiation can be different from a radiation fraction to another as a radiotherapy treatment is often divided into several fractions (30 for example for glioblastoma), typically spaced 24 hours apart, and models restricted to lymphocytes circulating in the blood may therefore significantly misestimate the lymphocyte pool dose.

[0006] As a result, none of the previously published models has yet considered the out-of-field (OOF) dose received by regions rich in lymphoid structures, nor the slow recirculation and homing of lymphocytes between lymphoid organs such as the spleen or lymph nodes and blood (which only contains about 2% of the total lymphocyte pool at any given time

[0020] ,

[0021] ). For these reasons, it is not trivial to extrapolate the doses received by circulating blood to the doses received by recirculating lymphocytes.

[0007] DETAILLED DESCRIPTION OF THE INVENTION

[0008] In this context, the present inventors propose a new in silica model with highly efficient numerical implementation based on stochastic continuous time semi-Markov chain process to evaluate doses received by the lymphocyte pool after a radiotherapy treatment, in particular after conventional dose rate (CONV) Volumetric Modulated Arc Therapy (VMAT) irradiation in patients with primary brain tumors. For the first time, they propose to take into account in this evaluation the slow recirculation and homing of lymphocytes between lymphoid organs such as the spleen or lymph nodes and blood and optionally an out-of-field low dose component (OOF) to lymphoid regions (e.g., spleen, lymph nodes, etc.) that are not within the irradiation zone.

[0009] More precisely, the model of the invention employs a continuous-time-Markov-chain simulation based on two interconnected compartment models representing lymphocyte recirculation and blood flow.

[0010] In the examples below, a 6MV photon irradiation plan with 30 fractions of 2Gy was considered, comparing conventional VMAT (l.lGy / min) to UHDR (lOOGy / s) dose rates. To assess the doses received by the lymphocyte pool during irradiation, their dynamics within the organs was modeled. The authors distinguished two distinct but interdependent processes: (i) the slow recirculation of lymphocytes between lymphoid organs, mediated by specific biological processes such as CD62L and SIP pathways for recirculation and homing

[0019] , and (ii) the rapid circulation of lymphocytes through all organs via the bloodstream. These 2 phenomena were modelled in this work by implementing 2 interconnected compartmental models which we will herein refer to as Mi and M2respectively. The dose received by a lymphocyte was randomly sampled from the brain dose-volume histogram when passing through the brain compartment. Additionally, an out-of-field low dose component (OOF) to the head and neck lymph nodes was in a first attempt modeled with a constant dose, delivering lOmGy to 30% of the lymph node compartment per 2Gy fraction. As indicated below, several scenarios were considered: an Hl scenario, which only considered blood circulation (M2model only), an H2 scenario taking into account lymphocyte recirculation (interconnected Mi and M2models) and an H3 scenario taking into account lymphocyte recirculation (interconnected Mi and M2models) as well as the out-of-field dose. At the end of treatment, CONV VMAT irradiation encompassed 100% of the blood, while UHDR irradiation covered only 30% with the Hl scenario. With the H2 scenario, CONV VMAT irradiated 32.8% of the lymphocytes with a mean dose of 25.6±27.0 mGy, whereas UHDR irradiation affected 2.4% of lymphocytes with a mean dose of 0.83±0.66Gy. In both cases (H2 - CONV and UHDR), the majority of irradiated lymphocytes passed through the treatment field only once. When considering the additional OOF low dose component in the H3 scenario, all lymphocytes were exposed to a low mean dose (approximately 50mGy) in both irradiation scenarios. Applying the in silica model of the invention, the present results thus show that i) UHDR reduced the fraction of lymphocytes irradiated compared to CONV VMAT, but with a higher average dose, without considering OOF and that ii) the contribution of the OOF dose to lymphoid structures should not be neglected to correctly model the dose received by the lymphocyte pool.

[0011] Based on this work and in particular the second observation, the present inventors propose a new method to assess the irradiation doses actually received by lymphocytes after a radiation therapy has been performed on a target organ of a patient, by considering one or two important features: ii) the slow recirculation and homing of lymphocytes between lymphoid organs such as the spleen or lymph nodes and blood, and i) optionally, the out-of-field (OOF) dose received by out-of-target regions rich in lymphoid structures (e.g., spleen, lymph nodes, etc.).

[0012] When the treatment plan, the dose map, and the structure delineation of the nearby lymphoid organs can be determined before the radiation treatment is performed on the patient, the method of the invention also enables to efficiently and reproducibly predict the radiation dose that will be received by the lymphocytes in said patient when he / she undergoes the planned treatment. In this case, it will therefore be possible to adapt the radiation plan and / or to identify an alternative treatment strategy that will better spare lymphocytes, to prevent or decrease severe radiation-induced lymphopenia (sRIL), so as ultimately to improve treatment outcomes.

[0013] The method of the invention is thus also useful for optimizing and / or personalizing a radiotherapy treatment plan to prevent radiation-induced lymphopenia to occur or to worsen. Although the present results have been obtained with a radiotherapy treatment performed on the brain, the model of the invention can be applied to other irradiation locations. From a clinical point of view, reducing the incidence of RIL could not only improve the response to radiotherapy and chemoradiation, but also enhance the potential of radioimmunotherapy combinations or abscopal effect.

[0014] In a first aspect, the present invention relates to a method for determining the irradiation dose that has been or will be received by a lymphocyte L (preferably by the lymphocytes pool) during a radiotherapy treatment performed on a target organ or location of a patient, said method comprising the steps of: a) collecting the following data associated to said radiotherapy treatment:

[0015] - the radiation treatment planning applied to or scheduled for said patient (Rl),

[0016] - the radiation reference doses applied to or scheduled for said patient (R2), and

[0017] - the number, location and volume of the lymphoid organs that have been or will be irradiated (if any, R3), b) simulating the lymphocyte path between the several organs of the body, c) determining the irradiation dose received by said lymphocyte L during said radiotherapy treatment would it be circulating in peripheral blood, or present in or circulating between lymphoid organs, d) optionally, determining the irradiation dose received by at least one out-of-field lymphoid organ during said radiotherapy treatment, e) combining these data in order to determine the irradiation dose that has been or will be received by the lymphocytes of said patient during said radiotherapy treatment.

[0018] Preferably, step d) is performed as this further permits to improve the irradiation dose determination of the lymphocytes pool in the patient. Therefore, the present invention also relates to a method for determining the irradiation dose that has been or will be received by a lymphocyte L during a radiotherapy treatment performed on a target organ of a patient, said method comprising the steps of a) collecting the following data associated to said radiotherapy treatment : - the radiation treatment planning applied to or scheduled for said patient (Rl),

[0019] - the radiation reference doses applied to or scheduled for said patient (R2), and

[0020] - the number, location and volume of the lymphoid organs that have been or will be irradiated (R3, if any), b) simulating the lymphocyte path between the several organs of the body, c) determining the irradiation dose received by said lymphocyte L during said radiotherapy treatment would it be circulating in peripheral blood, or present in or circulating between lymphoid organs, d) determining the irradiation dose received by at least one out-of-field lymphoid organ during said radiotherapy treatment, e) combining these data in order to determine the irradiation dose that has been or will be received by the lymphocytes of said patient during said radiotherapy treatment.

[0021] In the following, radiotherapy will be broadly defined as a treatment method using ionizing radiation. Generally speaking, this treatment can be divided into several NFfractions Ft( i e [1, JVF]) spaced by a period At (usually 24h for external radiotherapy). Each fraction Ftcan be made up of a succession of NAarcs or beams A; ,■ ( j e [l, tV^]), each lasting ^. .seconds and spaced Atmterjjea)nsecondsapart. Each arc can itself subdivided into Nssuccessive segments S;y fe(k e [1, TVs-]), with a duration

[0022] In the context of the invention, the said radiotherapy treatment can be an external radiotherapy in particular 3D conformal radiation therapy (3DCRT), intensity-modulated radiation therapy (IMRT), volumetric modulated arc therapy (VMAT), stereotactic body radiation therapy (SBRT), image-guided radiation therapy (IGRT), including MRI-associated radiotherapy, proton therapy, high dose rate radiation therapy (HDRT) or ultra-high dose rate (UHDR) or brachytherapy or internal radiotherapy.

[0023] In a preferred embodiment, said radiotherapy treatment is VMAT or UHDR. The target organ or location can thus be any organ or location that has been treated or is treatable by any of these radiotherapy treatments. In some cases, several organs or location can be treated simultaneously.

[0024] In a particular embodiment, as disclosed in the examples below, said target organ is the brain and said out-of-field lymphoid organs are the subcutaneous lymph nodes of the head-and-neck region.

[0025] In another particular embodiment, as disclosed in the examples below, the target organ is a thoracic organ, notably selected from lungs (including the left and right lungs), esophagus, mediastinum, pleura, or thymus. Thoracic organs rich in blood include heart, liver, "aorta" and "vena cava" (including the superior and inferior vena cava), and thoracic lymphoid organs including the spleen, bone marrow, and major thoracic lymph nodes. Major thoracic lymph nodes comprise right lymph nodes (including LI, L2, L3 right axillary, sus-clavicular and interpectoral right lymph nodes), left lymph nodes (including LI, L2, L3 left axillary, sus-clavicular and interpectoral left lymph nodes), mammalian chain lymph nodes, mediastinal lymph nodes and mesenteric lymph nodes. The specific in-field and out-of-field thoracic lymphoid organs will vary depending on the specifically targeted organ, the tumor location in this organ and the specific radiation treatment planning (RTP) and must be determined specifically for each patient, but the irradiation dose(s) received by at least one and preferably several or all out-of-field thoracic lymphoid organs is (are) determined. In particularly preferred embodiments:

[0026] • The target organ is right lung, radiation treatment planning (RTP) is such that in-field thoracic lymphoid organs comprise bone marrow, right lymph nodes (including LI, L2, L3 right axillary, sus-clavicular and interpectoral right lymph nodes), mammalian chain lymph nodes and mediastinal lymph nodes and out-of-field thoracic lymphoid organs comprise spleen, left lymph nodes (including LI, L2, L3 left axillary, sus-clavicular and interpectoral left lymph nodes) and mesenteric lymph nodes (see Figure 5, Pl), and the irradiation doses received by all out-of- field thoracic lymphoid organs are determined. Optionally, the irradiation dose(s) received by at least one and preferably several or all out-of-field non-thoracic lymphoid organs is (are) further determined;

[0027] • The target organ is left lung, radiation treatment planning (RTP) is such that in-field thoracic lymphoid organs comprise bone marrow, right lymph nodes (including LI, L2, L3 right axillary, sus-clavicular and interpectoral right lymph nodes), left lymph nodes (including LI, L2, L3 left axillary, sus-clavicular and interpectoral left lymph nodes), mammalian chain lymph nodes and mediastinal lymph nodes and out-of-field thoracic lymphoid organs comprise spleen and mesenteric lymph nodes (see Figure 5, P2) and the irradiation doses received by all out-of-field thoracic lymphoid organs are determined. Optionally, the irradiation dose(s) received by at least one and preferably several or all out-of-field non-thoracic lymphoid organs is (are) further determined;

[0028] • The target organ is esophagus, and radiation treatment planning (RTP) is such that in-field thoracic lymphoid organs comprise bone marrow, spleen, right axillary lymph nodes LI, L2, L3, left axillary lymph nodes LI, L2, L3, mammalian chain lymph nodes, mediastinal lymph nodes and out-of-field thoracic lymphoid organs comprise spleen, right sus-clavicular lymph nodes, right interpectoral lymph nodes, left sus-clavicular lymph nodes, left interpectoral lymph nodes and mesenteric lymph nodes (see Figure 5, P3), and the irradiation doses received by all out-of-field thoracic lymphoid organs are determined. Optionally, the irradiation dose(s) received by at least one and preferably several or all out-of-field non- thoracic lymphoid organs is (are) further determined.

[0029] Step a)

[0030] The following data associated to a radiotherapy treatment are to be gathered in order to implement the calculation of the method of the invention:

[0031] - the radiation treatment planning (RT planning) applied to or scheduled for said patient (Rl).

[0032] In radiotherapy, radiation treatment planning (RTP) is the process in which a team consisting of radiation oncologists, radiation therapists, medical physicists and medical dosimetrists plan the appropriate external beam radiotherapy or internal radiotherapy or brachytherapy treatment technique for a patient with cancer. Typically, medical imaging is used to form a virtual patient for a computer-aided design procedure. A CT scan is often the primary image set for treatment planning (in particular in external beam radiotherapy), while magnetic resonance imaging provides excellent secondary image set for soft tissue contouring. The two imaging modalities complemented by ultrasound imaging can be used as the primary imaging modality for preparing the brachytherapy plan. Positron emission tomography is less commonly used and reserved for cases where specific uptake studies can enhance planning target volume delineation. Modern treatment planning systems provide tools for multimodality image matching, also known as image coregistration or fusion. Treatment simulations are used to plan the geometric, radiological, and dosimetric aspects of the therapy using radiation transport simulations and optimization. For external RT, this process involves selecting the appropriate beam type (which may include photons, electrons and protons), energy (e.g. 6, 18 megavolts (MV) photons) and physical arrangements. In brachytherapy planning involves selecting the appropriate catheter positions and source dwell times (in High Dose Rate (HDR) or Pulsed Dose Rate (PDR) brachytherapy) or seed positions (in Low Dose Rate (LDR) brachytherapy). The radiation treatment planning applied to or scheduled for said patient (Rl) preferably comprises planning medical images, and more preferably a patient's external contour and a patient's anatomy (i.e. a three-dimensional representation with localization of all organs of the patient in the imaging area). Such planning medical images generally contain a limited field of view. Indeed, only partial medical imaging of the body, limited to the area of the target volume, is generally available in routine clinical context.

[0033] - the radiation reference doses applied to or scheduled for said patient (R2).

[0034] This can be done either manually or with computer-assisted technologies.

[0035] The radiation reference doses are however generally provided as irradiation dose maps (e.g. DICOM RTDose) of the whole treatment. The dose maps may have been obtained at the treatment planning stage or re-evaluated during treatment if images of the day's anatomy are available. Dose maps can be available or re-evaluated R2.1) for each irradiation fraction Fitor R2.2) for each arc or beam j, or R2.3) for each segments Sl Jik. Only one of R2.1, R2.2 or R2.1 is required.

[0036] The radiation reference doses are preferably provided as an in-field dose map from the treatmentplanning system.

[0037] - the number, location and volume of the lymphoid organs that have been or will be irradiated (R3, if any).

[0038] This can be done either manually or with computer-assisted technologies.

[0039] These data are however generally collected as a file (e.g a DICOM RTSS) describing the names and positions of the delineated structures that contains at least the annotation of the irradiated organs and the lymphoid structures in the irradiation field (and optionally further out of the irradiation field).

[0040] In the case of lymph nodes in the irradiation field, the fraction of lymph nodes taken into account in the irradiation field is also necessary and can be found in the literature (see Example).

[0041] Further data that may optionally be collected include: the gender of the patient (male or female) to personalize the blood circulation model used (01), and the names and positions of the delineated structures that contain at least the annotation of the irradiated organs and the lymphoid structures out of the irradiation field (02). These data can be further included in the file (e.g a DICOM RTSS) describing the names and positions of the delineated structures that contains at least the annotation of the irradiated organs and the lymphoid structures in the irradiation field.

[0042] In particular, in a preferred embodiment of the invention, it will be important to gather the following information:

[0043] 1) the gender of the patient, which will define which variant of the model (M2 male or female) to apply the blood circulation model (01), and

[0044] 2) the RTSS that contains at least the annotation of the irradiated organs and the lymphoid structures in (R3) and out (02) of the irradiation field. This can be done either manually or with computer-assisted technologies.

[0045] 3) his / her irradiation dose map (R2) along with its irradiation parameters (Rl, frequency and number of fractions; geometric and temporality of irradiation beams).

[0046] Knowing the contour of the target organ or vascularized region affected by irradiation, a dose-volume histogram (DVH) can then be generated for each segment / beam of the irradiation plan for in-field irradiated organ or vascularized region affected by irradiation.

[0047] 4) for lymph nodes, which are distributed throughout the body, the fraction of each category (subcutaneous, mesenteric, etc) that will be or have been irradiated (if any, also R3).

[0048] This can be done either manually or with computer-assisted technologies.

[0049] Step b) b) Simulation of the lymphocyte path between the several organs of the body

[0050] This step requires to model the lymphocytes dynamics within the organs. In this aim, two distinct but interdependent processes can be distinguished : (i) the slow recirculation of lymphocytes between lymphoid organs (e.g lymph nodes, spleen, Peyer's patches, bone marrow, lung ...), mediated by specific biological processes such as CD62L and SIP pathways for recirculation and homing

[0019] , and (ii) the rapid circulation of lymphocytes through all organs via the bloodstream. These 2 phenomena can be modelled by implementing 2 interconnected compartmental models referred to as and M2respectively.

[0051] In model M1 (each compartment represents a lymphoid organ (e.g lymph nodes, spleen, Peyer's patches, bone marrow, lung ...). For instance, the Model may be based on the Ganusov et al.

[0022] model of lymphocyte recirculation via circulating blood between lungs, liver, spleen, subcutaneous lymph nodes (SCLNs), mesenteric lymph nodes (MLNs) and Peyer's Patches (PPs), using murine experimental data. In this model, lung and liver compartments have extremely short mean residence times (0.46 min and 0.88 min respectively) compared with the other lymphoid compartments but captured most blood lymphocytes (78.2% and 17.4% respectively). This model can be modified by grouping 3 compartments {circulating blood, lung, liver} into a single compartment called Blood circulation. However, any alternative model Ml in which each compartment represents a lymphoid organ (e.g lymph nodes, spleen, Peyer's patches, bone marrow, lung ...) may be used, such as the models described in

[0023] ,

[0024] What is important for the method of the invention is that it contains a simulation of the slow recirculation of lymphocytes between lymphoid organs.

[0052] In particular, said step b) can be performed by using a model of lymphocyte recirculation via circulating blood between lungs, liver, spleen, subcutaneous lymph nodes, mesenteric lymph nodes, and Peyer's patches, by grouping the circulating blood, lung and liver in one unique compartment, and by determining the residence time of the lymphocytes in said compartment, as well as in spleen, lymph nodes, and Peyer's patches.

[0053] In accordance with

[0022] , residence times can be distributed according to an exponential law with a parameter the Spleen, PPs and Blood Circulation compartments, and according to a gamma distribution with a shape kM'1= 2 and a scale 0M1=Mean resic^-ence timef feorMl the NCLNs and MLNs compartments

[0054] In model M2, each compartment represents an organ of the body where the blood circulates. For instance, the Model M2may preferably be based on the ICRP Publication 89

[0025] , which describes a transition matrix for blood flow between 28 organs in the male and female human body.

[0055] Mean residence time values were adapted from the HEDOS model

[0026] , considering a total blood volume of 5.3 L and 3.9 L and a cardiac output of 6.5 L. min-1and 5.9 L. min-1for male and female patients respectively. In accordance with the data given in

[0012] , residence times were distributed

[0056] Mean residence time according to a Weibull distribution with a shape kM2= 2 and a scale AM2with r kM2 the gamma function. However, any alternative model M2 in which each compartment represents an organ of the body where the blood circulates may be used, such as the model describes in

[0026] ,

[0027] . What is important for the method of the invention is that it contains a simulation of the rapid circulation of lymphocytes through all organs via the bloodstream.

[0057] Finally, to simulate the path of a lymphocyte between compartments of a model (Ml or M2), a stochastic process can be used. In particular, a Markovian process or a semi Markovian process is preferably implemented. For instance, the stochastic process can be a continuous-time semi Markovian process.

[0058] A continuous-time semi-Markovian chain is a stochastic process with finite state space that changes states according to an arbitrary random variable and a transition matrix. In practice, the implementation of a continuous-time semi-Markov chain process consists of associating 2 parameters with each compartment of the Mi and M2models: 1) a continuous distribution of lymphocyte residence times, which enables a residence time to be stochastically determined for each lymphocyte passage through that compartment via, for instance, an inverse transform sampling (ITS), and 2) a vector of transition probabilities to the connected compartments, in order to stochastically determine which compartment the lymphocyte will migrate to.

[0059] More preferably, a stochastic process is implemented, in particular a continuous time Monte Carlo stochastic process, using, for instance, a Kinetic Monte Carlo algorithm, a dynamic Monte Carlo algorithm or a Gillespie algorithm.

[0060] However, alternative stochastic processes may also be used, such as discrete-time Markov chain, Markov jump process, hidden Markov chain, Bernoulli process, Harris chain, Wiener process or more generally random walk, martingales, Levy processes, Gaussian processes, random fields, renewal processes or branching processes.

[0061] Step c)

[0062] Step c) consists in determining the irradiation dose received by said lymphocyte L during said radiotherapy treatment would it be circulating in peripheral blood, or present in or circulating between lymphoid organs.

[0063] In a particular embodiment, step c) is performed by using any known method of irradiation dose calculation already proposed in the art, in particular by using the dose calculation proposed below in step e). Step d)

[0064] Step d) is optional although preferably present and consists in determining the irradiation dose received by at least one out-of-field lymphoid organ during said radiotherapy treatment.

[0065] In external beam radiotherapy (EBRT), non-zero doses are inevitably delivered outside the treatment field; this is often referred as out-of-field or peripheral dose. The 5% isodose is the most commonly used threshold to differentiate between in-field and out-of-field regions

[0028] ,

[0029] ,

[0030] ,

[0031] ,

[0032] ,

[0033] , The out-of-field doses are mostly low dose values (< 4 Gy), but the question of their potential impact on the probability of apparition of radio-induced cancers

[0034] and on the immune system is of the utmost importance. In particular, it has recently been shown that even very low doses deposited during imaging stages such as computed tomography (CT) scans contribute to the development of radiation- induced cancers

[0035] , In radiotherapy, this topic is currently experiencing renewed interest, particularly as modulated treatments (such as volumetric modulated arc therapy (VMAT) or intensity-modulated radiation therapy (IMRT)) are now routinely used in clinical care and are known to be associated with higher peripheral doses due to longer beam on times and irradiated volumes when compared to three- dimensional conformal radiotherapy (3D-CRT)

[0036] ,

[0037] ,

[0038] ,

[0039] , However, the results of recent studies remain mixed, with contradictory results depending on the populations studied

[0040] ,

[0041] ,

[0042] ,

[0043] ,

[0044] , In a very different context, the assessment of out-of-field doses seems increasingly crucial for optimizing radiotherapy treatments in the near future. Indeed, recent awareness of the immunomodulatory role of radiotherapy, coupled with the observation of a link between radiation- induced lymphopenia and patient response to treatment in several solid tumor sites, suggests the need to spare lymphocyte-rich structures as much as possible

[0045] , The fact that the lethal doses reported by several independent groups are of the order of a few Grays

[0046] ,

[0047] ,

[0048] is finally a strong argument demonstrating the clinical need for precise characterization of doses very close to the field.

[0066] Treatment planning systems (TPS) are used in clinical routine to estimate in-field dose distribution, but have been shown to systematically underestimate out-of-field dose for 3D conformational radiotherapy treatments, for intensity modulated treatments and for CyberKnife devices (Accuray, Sunnyvale, USA)

[0029] ,

[0036] ,

[0049] ,

[0050] ,

[0051] , Thus, despite clear clinical potential, the out-of-field dose computation is currently not available in clinical practice.

[0067] Two methods are currently used in the literature for out-of-field dose estimation for research purposes: Monte Carlo (MC) simulations, which are based on a stochastic approach and aim to estimate the average dose per voxel and its associated variance by simulating the tracking path of millions of incident particles knowing cross-sections of particle-matter interaction, and analytical approaches, which mathematically model the out-of-field dose, either based on physical or empirical hypotheses

[0052] , MC simulations and analytical methods can provide accurate out-of-field dose estimation

[0050] , but are today inappropriate for clinical routine implementation. On the one hand, MC simulations are hardware- and time-intensive, especially when it concerns the evaluation of low doses, and require detailed modeling of the irradiator, i.e. the linear accelerator (linac), which can be tedious when the blueprint of the device is not available. On the other hand, analytical models, which most of the time require new experimental measurements to adjust intrinsic parameters, are not suitable for large retrospective studies, which often incorporate irradiators (linacs) that are no longer available for experimental measurements.

[0068] Artificial intelligence and in particular technologies based on deep learning (DL) have drastically changed clinical practice in radiotherapy in recent years, allowing notably the automation of several time-consuming tasks, including segmentation

[0053] and treatment planning

[0054] , or the generation of virtual images (synthetic CT

[0055] ), demonstrating their ability to identify complex hierarchical features from spatially structured data

[0056] ,

[0069] In an exemplary embodiment, out-of-field dose can notably be decomposed in three main components: the patient scatter component, which corresponds to secondary photons resulting from a Compton interaction in the treatment field of primary photons, depositing their energy in the area outside the treatment field; the collimator scatter (or head scatter), component which results in doses deposited outside the treatment field by particles that have interacted with the collimator or other parts of the accelerator head, and the leakage component which is made up of primary particles that have not been intercepted by the accelerator jaws or other collimation parts.

[0070] The patient scatter component depends mainly on the size of the irradiated volume and the beam spectra, and is the majority component close to the field

[0057] , the two other components depend mostly on the geometrical properties of the linac used and its configuration during the treatment. The leakage component appears to be the dominant component far from the field

[0057] , and its amplitude is strongly dependent on the distance from the isocenter, as it depends on the angular shielding properties of the machine in a general way, i.e. integrating the attenuation properties of the jaws and the Multi-Leaf Collimator (MLC). While it is obvious that the information relating to the patient's scattering component is included in the in-field area, the absorbed doses generated by the head scattering component also generate signal in the field as the photons resulting from the Compton interaction are scattered throughout the volume

[0058] , The leakage component is ultimately a signature of the accelerator, making the task undoubtedly more complex for a neural network.

[0071] The method of the invention may use any of the approaches mentioned above: Monte Carlo (MC) simulations, analytical approaches, or artificial intelligence-based (including deep-learning, such as neural networks, and machine-learning) analysis as described in the examples below (see section II).

[0072] Artificial intelligence-based and in particular deep-learning and machine-learning analysis are particularly preferred as they do not involve the limitations in terms of computing time, difficulties in accessing experimental measurements and the lack of versatility inherent in analytical and MC methods (see Examples below, section II). However, MC simulations and analytical approaches may still be used alternatively.

[0073] In a preferred embodiment, data collected in step a) is directly or indirectly entered into a previously trained OOF dose prediction model (based on Monte Carlo simulations, analytical approaches, or artificial intelligence-based (including deep-learning, such as neural networks, and machine-learning), preferably an artificial intelligence-based OOF dose prediction model (e.g. a deeplearning or a machine-learning model), which returns the irradiation dose received by at least one out- of-field lymphoid organ during said radiotherapy treatment. When the OOF dose prediction model is a deep-learning OOF dose prediction model, it may preferably be based on one or more neural networks. Indeed, while the deep-learning OOF dose prediction model may rely only one neural network, it may also rely on two or more competing neural networks.

[0074] Preferably, the OOF dose prediction model has been trained (also referred to as "calibrated") on dose maps from a training population of cancer patients treated by radiotherapy. More preferably, the dose maps are whole body dose maps. The number of (whole body) dose maps from cancer patients treated by radiotherapy used for training the OOF dose prediction model is preferably at least 200, more preferably at least 250, at least 300, at least 400, even more preferably at least 500, at least 600, at least 700, at least 750, at least 800, at least 900, at least 1000, at least 1250, at least 1500, at least 1750 or even at least 2000. Indeed, the higher the number of (whole body) dose maps in the training data, the better will be the calculation of the OOF dose by the OOF dose prediction model for a new patient based on his / her data collected in step a).

[0075] The (whole body) dose maps may have been obtained from a single type of radiotherapy machine with the same beam characteristics (e.g. energy, presence of the flattening filter or not in photon treatments). In this case, the OOF dose prediction model will mainly be reliable for radiation treatment planning based on the type of radiotherapy machine and beam characteristics used for training.

[0076] In order to have an OOF dose prediction model reliable for a variety of radiotherapy machines and beam characteristics, the (whole body) dose maps have preferably been obtained from different radiotherapy machines (for instance selected from linear accelerators (linacs) from multiple manufacturers (e.g., Varian, Elekta, Accuray), older devices such as cobalt-60 units, betatron devices and, optionally, novel and emerging modalities such as ultra-high dose rate linacs used in FLASH radiotherapy, proton therapy systems, heavy ion therapy systems, and other advanced particle accelerators) with different beam characteristics, so that the model can then accurately predict OOF doses from several radiotherapy machines with several beam characteristics. The model may notably be trained on whole body dose maps from the French childhood cancer survivor study (FCCSS) cohort

[0059] used by the inventors, which comprises a high number of whole body dose maps obtained from various radiotherapy machines. In order for the OOF dose prediction model to reliably predict OOF doses for several types of radiotherapy machines with several beam characteristics, the number of (whole body) dose maps for each radiotherapy machine and beam characteristics setting is preferably at least 200, more preferably at least 250, at least 300, at least 400, even more preferably at least 500.

[0077] The whole-body dose maps needed to train the OOF dose prediction model may have been obtained using either analytical or Monte Carlo models.

[0078] In a particularly preferred embodiment, the OOF dose prediction model is an artificial intelligence-based OOF dose prediction model, such as a deep-learning or a machine-learning OOF dose prediction model, preferably a deep-learning OOF dose prediction model and more preferably a neural network-based (one or several competing neural networks) OOF dose prediction model, which has been trained on at least 500, more preferably at least 750, at least 1000, at least 1500 or even at least 2000 whole body dose maps from cancer patients treated by radiotherapy using different radiotherapy machines and beam characteristics.

[0079] In particular, when the OOF dose prediction model has been trained on whole body dose maps (preferred embodiment), the radiation treatment planning applied to or scheduled for said patient (Rl) comprises body medical imaging, and the radiation reference doses applied to or scheduled for said patient (R2) are provided as an in-field dose map from the treatment-planning system, the method according to the invention preferably further comprises before step d) a step dO) of generating a whole body dose map from: • the patient's external contour, and

[0080] • the in-field dose map from the treatment-planning system.

[0081] The in-field dose map from the treatment-planning system is directly available from data collected in step a).

[0082] The patient's external contour available from the medical imaging comprised in the radiation treatment planning (Rl) of the patient collected in step a) may be a whole body external contour (when whole body imaging has been performed, which is rare in clinical routine) or a partial external contour (when partial body imaging has been performed, which corresponds to most cases in clinical routine).

[0083] In the routine clinical context, most of the time, only a part of the patient's body has been imaged, and data collected in step a) includes an in-field dose map from the treatment-planning system and among other structures a partial patient's external contour and a partial patient's anatomy corresponding to the partial imaging zone.

[0084] In this case, the out-of-field lymphoid organ(s) for which an irradiation dose received during radiotherapy is determined in step d) may be limited to out-of-field lymphoid organ(s) present in the partial body medical imaging available from the treatment-planning system collected in step a). Then, step d) determines the irradiation dose received by at least one out-of-field lymphoid organ during said radiotherapy treatment based on the whole body dose map generated in step dO) and a partial patient's anatomy.

[0085] However, in a preferred embodiment, the out-of-field lymphoid organ(s) for which an irradiation dose received during radiotherapy is determined in step d) comprise not only out-of-field lymphoid organ(s) present in the partial body medical imaging available from the treatment-planning system collected in step a), but also out-of-field lymphoid organ(s) not present in the partial body medical imaging available from the treatment-planning system collected in step a).

[0086] When the radiation treatment planning applied to or scheduled for said patient (Rl) comprises partial body medical imaging (i.e. the patient's external contour and patient's anatomy available from the medical imaging comprised in the radiation treatment planning (Rl) of the patient collected in step a) are a partial external contour and a partial anatomy), step (dO) preferably comprises a sub-step (dOi) of determining a whole body external contour from the partial external contour, the whole body external contour then being used with the in-field dose map to generate a whole body dose map (Figure 6). The whole body external contour may be artificially generated from the partial external contour:

[0087] • either directly by the OOF dose prediction model (when the OOF dose prediction model has been trained to both determine a whole body external contour from a partial external contour and generate a whole body dose map from a patient's whole body external contour and an in-field dose map from the treatment-planning system, data collected in step a) (including a partial external contour and an in-field dose map) can then be directly entered into the OOF dose prediction model); or

[0088] • by a separate external contour extension model (when the OOF dose prediction model has been trained only to generate a whole body dose map from a patient's whole body external contour and an in-field dose map from the treatment-planning system). In this case, the infield dose map collected in step a) and the patient's whole body external contour obtained by external contour extension model from the partial external contour available from the radiation treatment planning (RTP) of the patient collected in step a) are then entered into the OOF dose prediction model. In this case, the external contour extension model may notably be based on, for instance, o 1) geometric shapes, where simple or complex forms (such as ellipsoids, cylinders, spline-based surfaces or others) are used to extrapolate segmentation to encompass the entire body or o 2) atlas-based methods wherein a pre-defined library of anatomical templates is registered to the patient's original image using rigid, affine or deformable registration techniques (such as homography), allowing for propagation of contours to regions not present in the field. o 3) Al-driven image extension methods (using for instance deep learning algorithms, such as convolutional neural networks or generative models) trained on a dataset composed of partial external contours associated with corresponding whole body external contours (for instance the contour used for the deep learning OOF dose prediction model training).

[0089] Step d) then determines the irradiation dose received by at least one out-of-field lymphoid organ during said radiotherapy treatment based on the whole body dose map generated in step dO) and a whole body patient's anatomy.

[0090] The patient's anatomy available from the medical imaging comprised in the radiation treatment planning (Rl) of the patient collected in step a) may be a whole body anatomy (when whole body imaging has been performed, which is rare in clinical routine) or a partial anatomy (when partial body imaging has been performed, which corresponds to most cases in clinical routine).

[0091] When the radiation treatment planning applied to or scheduled for said patient (Rl) comprises partial body medical imaging (i.e. the patient's anatomy available from the medical imaging comprised in the radiation treatment planning (Rl) of the patient collected in step a) is partial anatomy), a whole body patient's anatomy may be artificially generated from partial anatomy using any method disclosed above for generating a whole body external contour from a partial external contour, based on training data comprising partial and whole body anatomy for the same patient.

[0092] Therefore, in a preferred embodiment:

[0093] • step d) is performed by directly or indirectly entering data collected in step a) into a previously trained OOF dose prediction model, wherein the OOF dose prediction model is an artificial intelligence-based OOF dose prediction model, such as a deep-learning or a machine-learning OOF dose prediction model, preferably a deep-learning OOF dose prediction model and more preferably a neural network-based (one or several competing neural networks) OOF dose prediction model, which has been trained on at least 500, more preferably at least 750, at least 1000, at least 1500 or even at least 2000 whole body dose maps from cancer patients treated by radiotherapy using different radiotherapy machines with different beam characteristics (preferably, the number of whole body dose maps for each radiotherapy machine and beam characteristics setting is preferably at least 200, more preferably at least 250, at least 300, at least 400, even more preferably at least 500),

[0094] • the radiation treatment planning applied to or scheduled for said patient (Rl) comprises partial body medical imaging,

[0095] • the radiation reference doses applied to or scheduled for said patient (R2) are provided as an in-field dose map from the treatment-planning system,

[0096] • the method further comprises before step d) a step dO) generating a whole body dose map from: o the patient's external contour, available from the radiation treatment planning (RTP), and o the in-field dose map from the treatment-planning system, and

[0097] • Step d) determines the irradiation dose received by at least one out-of-field lymphoid organ during said radiotherapy treatment based on the whole body dose map generated in step dO) and a whole body patient's anatomy. In a particular embodiment, when applied, said step d) is performed on at least one, preferably on all, out-of-field lymphoid organ chosen from spleen, subcutaneous lymph nodes, mesenteric lymph nodes, and Peyer's patches via analytical, Monte-Carlo or artificial intelligence-based methods, preferably via artificial intelligence-based methods, and in particular via deep-learning and machine learning methods (see section II of Examples).

[0098] Step e)

[0099] Step e) consists in combining the data obtained in the previous steps in order to determine the irradiation dose that has been or will be received by the lymphocyte population of said patient during said radiotherapy treatment.

[0100] Knowing the information described in step a), the lymphocyte path between organs generated in step b), the doses measured in step c) and optionally but preferably the out-of-field dose map generated in step d), it is possible to evaluate the dose received by the lymphocyte during the treatment.

[0101] To do that, one reasoning can be: a given lymphocyte circulating in Mi is likely to be irradiated if 1) it is in an in field irradiated lymphoid organs of Mi model during beam-on-time or 2) it is in the M2Blood circulation compartment during the beam-on time or optionally but preferably 3) it is in an out-of-field irradiated organ of Mi model during beam-on-time. For each of these situations, the dose assessment methods are described below:

[0102] Case 1)

[0103] Let L be a given lymphocyte and 8 a time interval during which L is in a Mi irradiated compartment. L receives a non-zero dose if and only if there exists at least one irradiation segment Stjkactive during the time interval A(S;fe) such that A(S;J fe) A 8 #= 0. Noting = [ Sijk\ &(Sij'k) A 8 #= 0 } the set of segments that irradiate lymphocyte L during its passage through the M2Brain during the time interval 8, the dose Dg received by L is evaluated stochastically via an inverse transform sampling as Dg = with Ui j ka random variable following a continuous uniform distribution between 0 and 1: Uijk~ tf[o,i] and || || the time interval duration. Case 2)

[0104] The lymphocyte first enters M2via the Large Veins. The dose estimation process is then the same as in case 1), by replacing "M / ' by "M2"

[0105] Case 3) (optional but preferred)

[0106] With the out-of-field dose map optionally generated at optional step d), along with the contour data of out-of-field lymphoid structure optionally collected in the step a), it is possible to generate irradiation DVH for lymphoid organs outside the irradiation field. With them, it is possible to estimate the dose received by a lymphocyte in this organ during the irradiation. The dose estimation process is then the same as in case 1).

[0107] Finally, the total dose received by the lymphocytes is the sum of all the contribution from cases 1), 2) and 3). The same process is then performed on multiple lymphocytes (105for example) to estimate stochastically the dose received by the lymphocyte pool.

[0108] Preferably, said step e) enables to calculate the percentage of lymphocytes that have been or that will be lethally irradiated in said patient during said radiotherapy treatment.

[0109] Said step e) can also help to determine a dose volume histogram of the lymphocyte pool.

[0110] The whole method of the invention, and more preferably the combination step e), is advantageously implemented by a computer.

[0111] The said method can be used after treatment planning / treatment; or at the treatment planning / replanning stage to implement a personalized lymphocyte sparing strategy.

[0112] The said method can be used for optimizing and / or personalizing a radiotherapy treatment plan to prevent radiation-induced lymphopenia.

[0113] The said method can be used for selecting a radiotherapy treatment and / or optimizing prescription in order to minimize the detrimental effects of radiotherapy onto the immune system of said patient.

[0114] The present invention also relates to the use of the irradiation dose received by at least one out-of-field lymphoid organ during a radiotherapy treatment, in order to optimize and / or personalize said radiotherapy treatment to prevent radiation-induced lymphopenia, preferably for implementing the method according to the invention. The present invention also relates to a computer device configurated to implement the method according to the invention.

[0115] FIGURE LEGENDS

[0116] Figure 1: A, compartment models Mi and M2describing lymphocyte recirculation between lymphoid organs and blood flow, respectively. B, a schematic representation of a recirculating lymphocyte path and the dose it receives in the continuous time semi-Markov chain simulation. The figure presents the case of a CONV irradiation of a brain lesion consisting of 30 fractions of 3 arcs (VMAT irradiation). For variant Hl, only the M2model is considered: the lymphocyte does not recirculate outside the blood (prior art method simulating only blood circulation of lymphocytes using model M2). For variant H2, a lymphocyte first circulates in model Mp If it is in compartment Mp Blood circulation during an irradiation fraction, it enters and circulates in model M2. It is effectively irradiated if it is in M2Brain compartment when the beam is active (hatched red and white). The dose it receives, written in a simplified form in the figure, is proportional to the time Atsspent in each irradiation segment and is then evaluated stochastically via Inverse Transform Sampling from the whole-brain Dose Volume Histogram. Variant H2 corresponds to the method according to the invention without optional step d). With variant H3, the lymphocyte can also be irradiated out-of-field, in SCLNs in the head and neck region. Variant H3 corresponds to the method according to the invention with step d). SCLNs: subcutaneous lymph nodes, MLNs: mesenteric lymph nodes, PPS: Peyer's Patches (PPs), OOF: Out-of- field

[0117] Figure 2: Three variants of the model were tested for a CONV irradiation of a brain lesion consisting of 30 fractions of 3 arcs (VMAT irradiation): Hl / panel A) lymphocytes don't recirculate out of the blood i.e., lymphocytes circulate in M2only (prior art method); H2 / panel B) lymphocytes recirculate between lymphoid organs i.e. Mp and M2interconnected (method according to the invention without optional step d); H3 / panel C) lymphocytes recirculate between lymphoid organs and can be irradiated with OOF dose in the H&N lymphoid structures (method according to the invention with step d). The upper graph represents Dose Volume Histograms (DVH) of blood (Hl) and of the recirculating lymphocyte pool without (H2) and with (H3) consideration of the low out-of-field (OOF) dose in the lymph nodes of the H&N region at the end of CONV VMAT treatment. In each figure, the mean DVH over the whole cohort is shown in solid dark line, the associated standard deviation in dotted line and the DVHs for each patient in the cohort in solid light grey. For H2 DVH, a zoom subplot is also shown. On the lower graphs, the average number of lymphocyte irradiations (left), and the associated dose (right), averaged over the whole cohort are described. For model H3, lymphocytes can be irradiated either in or out of the field, and the average contribution of each type of irradiation is shown.

[0118] Figure 3: Dose / Volume histograms to the blood (upper panel, model M2only, no recirculation) and to the recirculating lymphocytes (model Mi and M2interconnected), with (lower panel) or without (middle panel) the out-of-field dose component to head and neck lymph nodes at the end of the treatment (30 fractions) for 2 different dose rates (conventional VMAT treatment: dark grey line or UHDR light grey line). Simulation performed on 105lymphocytes.

[0119] Figure 4: The three variants of the model were tested as in figure 2, but with a UHDR treatment. The upper graph represents Dose Volume Histograms (DVH) of blood (Hl / panel A) and of the recirculating lymphocyte pool without (H2 / panel B) and with (H3 / panel C) consideration of the low out-of-field (OOF) dose in the lymph nodes of the H&N region at the end of the UHDR treatment. In each figure, the mean DVH over the whole cohort is shown in solid dark line, the associated standard deviation in dotted line and the DVHs for each patient in the cohort in solid light grey. For H2 DVH, a zoom subplot is also shown. On the lower graphs, the average number of lymphocyte irradiations (left), and the associated dose (right), averaged over the whole cohort are described. For model H3, lymphocytes can be irradiated either in or out of the field, and the average contribution of each type of irradiation is shown.

[0120] Figure 5: Contours of thoracic lymphoid organs (including the spleen and major thoracic lymph nodes. Major thoracic lymph nodes comprise right lymph nodes (including LI, L2, L3 right axillary, supraclavicular, and interpectoral right lymph nodes), left lymph nodes (including LI, L2, L3 left axillary, supraclavicular, and interpectoral left lymph nodes), mammary chain lymph nodes, mediastinal lymph nodes, mesenteric lymph nodes and bone marrow,) in the 3 patients Pl, P2, and P3, as well as the 5% isodose delineating the "in-field" region from the "out-of-field" region.

[0121] Figure 6: Illustration, in the case of thoracic irradiation, of contour extension strategies to evaluate the doses received by lymphoid organs outside the irradiation field. In most cases, in clinical practice, partial body imaging is performed, and it is then possible to calculate a "partial body out-of-field dose" in the imaged part of the body (thorax here). It is also possible to extend the patient's contour to the whole body in order to calculate a "whole body out-of-field dose map". When a full body image of the patient is taken during clinical routine, and it is then possible to calculate the whole body out-of-field dose map directly, without the need to extend the anatomy.

[0122] Figure 7: DVHs associated to segmented structures considered in LymphoDose, for the 3 patients treated with thoracic irradiations. These structures are divided into 2 categories: blood containing organs, used in the LymphoDose Hl and H3 models, and lymphoid organs, used in the H3 model only. The contours are those of patient P2. DVHs of the aorta and vena cava were corrected to correspond to the M2 compartments aorta and large arteries and large veins

[0123] Figure 8: DVH of blood (Hl) and of recirculating lymphocytes, considering the doses to lymphoid organs in the irradiation field (H3), at the end of treatment for the three patients studied

[0124] Figure 9: For the same 3 patients, the EDIC and EDIC_Xu scores, and the mean doses from LymphoDose Hl and H3.

[0125] EXAMPLES

[0126] / . Method of the invention

[0127] Materials and methods

[0128] Modeling lym phocyte pool dynam ics

[0129] To assess the doses received by the lymphocyte pool during irradiation, their dynamics within the organs was modeled. We distinguished two distinct but interdependent processes: (i) the slow recirculation of lymphocytes between lymphoid organs, mediated by specific biological processes such as CD62L and SIP pathways for recirculation and homing

[0019] , and (ii) the rapid circulation of lymphocytes through all organs via the bloodstream. These 2 phenomena were modelled in this work by implementing 2 interconnected compartmental models which we will herein refer to as Mi and M2respectively. As illustrated in Figure 1, both models implement a stochastic continuous-time semiMarkov chain process

[0060] to simulate lymphocyte paths between compartments during the whole treatment. A continuous-time semi-Markov chain is a stochastic process with finite state space that changes states according to an arbitrary random variable and a transition matrix. In practice, the implementation of a continuous-time semi-Markov chain process consists of associating 2 parameters with each compartment of the Mi and M2models: 1) a continuous distribution of lymphocyte residence times, which enables a residence time to be stochastically determined for each lymphocyte passage through that compartment via inverse transform sampling (ITS), and 2) a vector of transition probabilities to the connected compartments, in order to stochastically determine which compartment the lymphocyte will migrate to.

[0130] Model Mi (Figure 1) is based on the work of Ganusov et al.

[0022] , who proposed a model of lymphocyte recirculation via circulating blood between lungs, liver, spleen, subcutaneous lymph nodes (SCLNs), mesenteric lymph nodes (MLNs) and Peyer's Patches (PPs), using murine experimental data. In this model, lung and liver compartments have extremely short mean residence times (0.46 min and 0.88 min respectively) compared with the other lymphoid compartments but captured most blood lymphocytes (78.2% and 17.4% respectively). Based on the hypothesis that lymphocytes circulating in these 2 organs do not a priori leave the bloodstream, the model was modified in the present study by grouping 3 compartments {circulating blood, lung, liver} into a single compartment called Blood circulation. In accordance with

[0022] , residence times were distributed according to an exponential law with a parameter = - for the Spleen, PPs and Blood Circulation compartments,

[0131] Mean residence time and according to a gamma distribution with a shape kM'1= 2 and a scale 0M1=Mean resic^-ence timeforfeMl the NCLNs and MLNs compartments.

[0132] The parameters used in the present example are:

[0133] Model M2(Figure 1) was based on the ICRP Publication 89

[0025] , which describes a transition matrix for blood flow between 28 organs in the male and female human body. Mean residence time values were adapted from the HEDOS model

[0026] , considering a total blood volume of 5.3 L and 3.9 L and a cardiac output of 6.5 L. min-1and 5.9 L. min-1for male and female patients respectively. In accordance with the data given in

[0012] , residence times were distributed according to a Weibull

[0134] Mean residence time distribution with a shape kM2= 2 and a scale AM2, with r the gamma function. kM2

[0135] The parameters used for male individuals in the present example are:

[0136]

[0137]

[0138] The parameters used for female individuals in this example are

[0139] T1

[0140] The code was developed in Julia language version 1.9.3

[0061] and simulations were performed considering 105particles.

[0141] Simu lation of patients with gl ioblastoma treatment

[0142] A cohort of patients with glioblastoma treated after 2018 with CONV VMAT irradiation was created. To simulate a treatment, each patient was characterized by 1) his gender, which defined which variant of the M2model (male or female) to apply, and 2) his irradiation parameters. For a given patient, a CONV VMAT treatment was defined by a succession of NFfractions ( i e [1, JVF]) spaced 24 hours apart. Each fraction Ftwas made up of a succession of NAarcs Aij ( j e [1, ^]), each lasting tAI., J seconds and spaced 5 seconds apart. Each arc At>J , was itself subdivided into Ns successive segments tAi■

[0143] $i i k (^6[1> Ns]), with the assumption that its duration was ts= NAx Ns dose maps associated with each segment were recalculated from the clinical treatment plan (CONV) with a Collapsed-Cone Convolution dose engine provided by TheraPanacea company (Paris, France). Using the brain contour, a whole-brain volume-dose histogram vdhijkwas then generated for each segment.

[0144] In order to evaluate the doses received by lymphocytes using the models described above, the logic is as follows (Figure 1): since the brain contains few lymphoid organs, a given lymphocyte circulating in Mi is likely to be irradiated in the field if it is in the Mi Blood circulation compartment during the beam-on time. In this case, the lymphocyte enters M2via the Large Veins. The latter is then effectively irradiated in the field if it is in the Brain during the beam-on time. Let L be a given lymphocyte and 6 a time interval during which L is in M2Brain. L receives a non-zero dose if and only if there exists at least one irradiation segment S;y feactive during the time interval A(S;J fe) such that A(Sij fe) n 6 #= 0. Noting Fig = { Si ] k| A(Si j k) A 5 =£ 0 } the set of segments that irradiate lymphocyte L during its passage through the M2Brain during the time interval 8, the dose Dg received by L is evaluated stochastically via an inverse transform sampling as Dg = 0 ng which L is in the M2Brain compartment, the total dose received by L during treatment is D = Sa e o Ds. Out-of-field dose evaluation

[0145] Because of the high radiosensitivity of lymphocytes

[0015] , the contribution of low doses outside the brain irradiation field and delivered to the H&N region, rich in lymphoid structures, was also considered. To do this, the right and left lymph nodes areas, including la, lb, II, III, IVa, V, Vila et VI lb regions

[0062] , were automatically contoured using ART-Plan software vl.11.5 (Therapanacea, Paris, France). As Treatment Planning Systems (TPS) used in clinical practice are known for their inaccurate evaluation of low doses (

[0052] ;

[0030] ), an in-house deep learning neural network (NN) was used to estimate the OOF dose to H&N lymphoid structures (cf. Example II below for evaluating OOF doses generated by photon external beam radiotherapy).

[0146] This in-house NN has been trained on more than 2,200 whole body dose maps from patients of the French childhood cancer survivor study (FCCSS) cohort

[0059] treated with 18 different machines between 1945 and 2011. As shown below, it is a universal model for evaluation of OOF doses generated by photon external beam radiotherapy (linear accelerator of high voltage > 1MV and60Co units) and can generate the "whole body" dose map from the patient's external contour and the in-field dose map from the TPS. Together, lymphoid structures contour and H&N OOF dose map allow calculation of the average OOF dose to lymphoid structures in the H&N region for each patient: OOF = —-, with v

[0147] Zu Vitand dtthe volume and the mean dose received by the lymphoid structure i, respectively. A Pearson test was performed to evaluate the correlation between OOF dose and irradiation parameters. Finally, based on the analysis of 1,200 lymph nodes performed on the Visible Human Data Set

[0065] , having shown that 31.5 ± 0.7% of subcutaneous lymph nodes (simply defined as non-mesenteric as in

[0022] ) are located in the H&N region, a lymphocyte had a probability of 0.315 of receiving an OOF dose if it is in the NCLNs during an irradiation fraction.

[0148] Model com parison and sensitivity analysis

[0149] To compare different physical and biological assumptions, the dose to lymphocytes has been computed with three variants of the model: Hl) circulation only in the bloodstream i.e., lymphocytes circulate in M2only; H2) with recirculation between lymphoid organs i.e. Mi and M2interconnected; H3) with recirculation between lymphoid organs and with OOF dose to H&N lymphoid structures. Finally, a sensitivity analysis was carried out to assess the impact of the uncertainties impacting the model (H2 and H3) parameters on the doses and irradiation volumes predicted for lymphocytes: for one representative patient (Patient 3), a variation of ± 50% on the mean residence times of the Mi and M2compartments and on the percentage of H&N lymph nodes receiving OOF dose was therefore performed. In addition, as the machines used in the NN training cohort to assess OOF doses were relatively outdated and probably gave higher OOF doses than current machines, a variation of ± 50% in OOF values was also applied. Results

[0150] Lymphocyte dynam ics

[0151] The blood and lymphocyte distributions between compartments as well as mean transit times and mean return times of Mi and M2compartments were evaluated. In particular, compartment Mi Blood circulation (grouping 3 compartments circulating blood, lung and liver) contained on average 6.38 ± 0.25 % of lymphocytes. Its mean transit time was estimated at 25.2 min and its mean return time was equal to 6.1 ± 7.2 hours. Compartment M2Brain contained on average 1.82 ± 2.12 % of blood (male and female), with a mean transit time of 4.9 ± 2.6 s (resp. 4.0 ± 2.1 s) and a mean return time of 6.7 ± 6.9 min (resp. 5.4 ± 6.9 min) for male (resp. female).

[0152] Cohort of patients with gl ioblastoma The cohort consisted of 33 patients with glioblastoma, whose characteristics and treatment parameters are described in Table 1. The cohort was composed of 20 male and 13 female patients with a prescribed dose of 60Gy, 50Gy and 40.5Gy for 25, 5 and 3 patients respectively. The planning treatment volumes (PTVs) had a mean volume of 278 ± 121 cm3and were irradiated at a mean dose rate of 1.14 ± 0.47 Gy / min.

[0153] Patients with glioblastoma (total = 33)

[0154] Prescribed dose (Gy) 60 25 (75.8%)

[0155] 50 5 (15.1%) 40.5 3 (9.1%)

[0156] Table 1 : Table describing the cohort of 33 patients with glioblastoma, std: standard deviation

[0157] Out-of-field dose

[0158] The average CONV treatment OOF dose to lymph nodes of the H&N for the whole cohort was 1.94 ± 0.36 Gy, with a maximum of 2.67Gy (for a prescribed dose of 60Gy to 384 cm3temporal PTV) and a minimum of 1.1 Gy (for a prescribed dose of 60Gy to 101 cm3fronto-temporal PTV). OOF values were correlated with PTV volume (r = 0.45, p = .009), which indicates that large volumes of irradiation generate more OOF dose

[0066] ,

[0159] Estimated doses to blood and lymphocytes

[0160] Variant Hl (Figure 2, Hl) resulted in the irradiation of almost all the circulating blood (V>OGV= 99.8 + 0.7%) at the end of VMAT treatment, at a mean dose of 309.9 ± 74.7 mGy (Fig. 1). The irradiated lymphocytes pass through the field at an average of 10.0 ± 4.9 times during the NFtreatment fractions (Figure 2, Hl). Considering lymphocyte recirculation in H2 variant (Table 2 and Figure 2, H2), the fraction of lymphocytes receiving a non-zero dose dropped to 40.4 ± 10.2%, with a lower mean dose estimated at 52.6 ± 21.1 mGy, especially because irradiated lymphocytes passed through the irradiation field only 1.58 ± 0.91 times on average. The most realistic scenario, including the OOF dose component (variant H3) (Table 2 and Figure 2, H3), showed that almost all lymphocytes (97.6 + 2.5%) were irradiated (in the field or OOF in the H&N lymph nodes) at the end of the VMAT treatment, mostly due to the OOF dose (mean dose equal to 265.6 ± 48.5 mGy). Model Average dose D50% D2% (mGy) (mGy) (mGy)

[0161] Table 2: Different metrics summarizing the volume of lymphocytes irradiated and the dose received by the blood (Hl) and of the recirculating lymphocyte (H2 and H3) during CONV VMAT irradiation, averaged over the whole cohort. V>0Gy(volume of lymphocytes having received a non-null dose), V>0125Cy(volume of lymphocytes having received a dose greater or equal to 0.125Gy), the average, median and 98th percentile doses are reported. The last 3 metrics are calculated on all blood / lymphocytes (with or without irradiation) and on irradiated blood / lymphocytes only (V>0Gy). For H3, V>0Gyrepresents lymphocytes that have been irradiated at least once either in the field or out of the field in the nodes of the H&N region.

[0162] Sensitivity analysis

[0163] Sensitivity analysis (Supplementary Data 5) revealed that the H3 CONV model is particularly sensitive to the fraction of lymph nodes in the H&N region exposed to OOF dose, where a variation of +50% (resp. - 50%) induced a variation in mean lymphocytes dose of +32% (resp. -28%). Moreover, a reduction of 50% of Mi SCLN mean transit time reduced the irradiated fraction by 10% (V>oGy from98% to 88%) and the mean dose by 43% (to 166 mGy). Finally, a 50% reduction in the OOF dose reduced the mean dose of 46% (to 154mGy) without changing the irradiated fraction.

[0164] Discussion

[0165] This model is the first to simulate radiation doses received by lymphocytes during standard of care brain irradiation considering (i) that lymphocytes are mostly found in lymphoid organs and not only in the peripheral circulation and (ii) that out of the field doses may significantly contribute to radioinduced lymphopenia. We used two interconnected compartmental models that mimic the presumed two-speed average journey of lymphocytes between blood, where they migrate fast, and lymphoid organs, where they may reside for a long time. Considering in- and out-field irradiation, the simulation revealed that 97.6% of total body lymphocytes may have passed through a radiation-receiving area during the full course of radiation therapy in patients treated with chemoradiation for glioblastoma, with 82.2% ± 9.5% of total body lymphocytes receiving doses >0.125 Gy. On average on the patient cohort, irradiated lymphocytes would receive a mean cumulative radiation dose of 265.6 mGy ± 48.5 mGy when treated with brain-directed conventional VMAT for glioblastoma.

[0166] Other teams previously intended to simulate radiation doses received by circulating lymphocytes upon brain irradiation, without considering recirculation or OOF doses

[0012] ,

[0013] , When considering that all lymphocytes are kept in the peripheral circulation as a closed system during radiation treatment of glioblastoma, our Hl model estimated that 99.8% ± 0.7% of all circulating lymphocytes received radiation at any dose during treatment, with a mean dose of 309.9 mGy ± 74.7 mGy. This is consistent with the findings obtained by Shin et al. who modelled a dose to circulating blood cells after 30 fractions of brain-directed Intensity Modulated Radiation Therapy (IMRT) treatment

[0021] , Hammi et al. obtained a blood mean dose of 133 mGy after 30 fractions of brain-directed IMRT when precisely modelling the cerebral vasculature using a sophisticated 4D blood flow model

[0013] , However, accurate estimation of the radiation dose delivered to lymphocytes is a complex task considering lymphocytes' dynamic within the body. Lymphocytes can traffic along lymph vessels and through lymphoid organs, thus transiently exit the blood circulation and reintegrate it later, implying that the whole blood-circulating lymphocyte pool can be renewed up to 11 times a day (31,32). Thus, peripheral lymphocytes present in the blood at the time of irradiation can be different from a radiation fraction to another, and models restricted to circulating lymphocytes may significantly misestimate the lymphocyte dose.

[0167] Thus, considering these recirculation mechanisms within the H2 and H3 variants, the lymphocyte dose values estimated by our model (Figure 2, H2 and H3) can be compared with the radiosensitivity of lymphocytes. A recently published review by Paganetti

[0015] synthesized multiple several in vivo studies and estimated that the average T lymphocytes dose-response a values was ~0.6Gy -1, which corresponds to a 50% lethal dose (LD50) of 1.15Gy with a linear model of cell survival. Moreover, lymphocytes radiosensitivity has been shown to depend on multiple factors, including subtype lineage (B cells being the most sensitive, followed by T cells and NK cells), activation status (naive cells being more sensitive than educated ones), function and most importantly, local microenvironment at the time of radiation, with tumor-infiltrating lymphocytes being more radioresistant than circulating ones

[0069] ,

[0070] ,

[0071] ,

[0072] , II. Deep-learning for instantaneous estimation of the out-of-field dose

[0168] Material and methods

[0169] Description of the dataset

[0170] The French Childhood Cancer Survivor Study (FCCSS) cohort, whose primary aim is to study long-term effects of children and adolescents treated for cancer, has been used in this work. More than 7000 patients under 21 years of age treated in 5 French centers between 1945 and 2000 for solid cancer or lymphoma composed this cohort. With the goal to develop a deep learning model for out-of-field dose estimation suitable for megavoltage photon irradiations, the following inclusion criteria were considered: 1) patients treated with a photon beam 2) treatments using linear accelerators with a high voltage > 1 MV or 60Co irradiators. 3310 patients were selected at this stage. For all patients in the cohort, whole body dose map was reconstructed, using an analytical method originally developed for bone marrow dose analysis

[0073] , This empirical method concatenated the 3D dose map as estimated by the Isogray TPS (Dosisoft, Cachan, France) in in-field areas; each treatment plan having being resimulated by an experienced operator based on the treatment details; to an out-of-field dose estimation obtained device-wise. For the last stage, the values for off-axis-ratios were derived from inwater phantom measurements performed and gathered since the mid-1980s (

[0074] ). All 3D whole body dose maps had voxels dimensions of 2x2x2 mm3. The exclusion criteria applied to these whole-body dose maps were as follows: 1) outlier dose maps, i.e. maps presenting local doses higher than 100 Gy, which had no physical justification in view of the doses prescribed (N = 104) were deleted (Radiation Dose in Radiotherapy from Prescription to Delivery, 1996), 2) corrupted dose maps (N = 2) were removed. At the end, data from 3204 patients were kept for this study. Table 1 summarizes the distribution by center of these patients. Twenty-five accelerators, here grouped into 3 categories: standard linacs, cobalt units and betatron units, were considered. The subset included 38 different pathologies, of which the most represented were nephroblastoma and other nonepithelial renal tumors (695 patients), Hodgkin lymphomas (449 patients) and astrocytomas (235 patients). Table 1: Characteristics of the patients selected in this work from the FCCSS cohort described at a center scale.

[0171] Data preprocessing

[0172] Several preprocessing steps were applied to the analytical dose maps to make them deeplearning compliant. These included: 1) padding into [370, 242, 1131] matrix sizes, 2) resampling of padded SAS files to [128, 128, 512] sizes, 3) extraction of in-field and out-of-field dose maps from whole-body dose maps (a 5% isodose threshold was chosen in this goal considering the maximum dose per patient as the reference dose), 4) creation of binary masks from the whole-body dose maps by a thresholding method separating the background from the foreground. An on-the-fly preprocessing pipeline was then applied using Medical Open Network for Artificial Intelligence (MONAI 0.8.0)

[0074] , including in the following order: loading, normalization, resampling, and concatenation. During the normalization step, 3D dose maps intensities were normalized by 100 Gy, to provide the neural network with values within [0,1], The on-the-fly resampling step was implemented to test the impact of batch size as a function of available VRAM (video random-access memory). A nearest neighbor interpolation strategy was used for the resampling step. Finally, the last step concatenated the whole-body dose maps and the in-field dose maps to provide input data of dimension Bx2xHxWxD to the neural network (with B, H, W and D respectively the size of the batch during the training process, the height, the width and the depth of the matrix). Figure 1 summarizes the preprocessing pipeline. Neural network training

[0173] A conventional 3D U-Net (

[0075] ;

[0076] ;

[0077] ), composed of four down-sampling blocks followed by four up-sampling blocks, was implemented. The Mean Square Error (MSE, see Equation 1) evaluated only on the foreground voxels outside the 5% isodose, i.e. only in the region considered in this paper as the out-of-field dose, was selected as loss function.

[0174] With D_(nn,i) and D_(gt,i) the normalized doses to the voxel i estimated respectively with the neural network and from the ground truth analytical method, and n the number of voxels considered.

[0175] Learning rate and weight decay were considered in the ranges [le-7,le-3] and [le-8,le-4], respectively, with le-4 and le-6 providing the optimal results. The Adam optimizer was used to update the network parameters. Instance normalization was preferred. Batch size of 20 corresponding to a resampled size of 64x64x256 was selected. The 3D U-Net was trained for 600 epochs (~50 hours) on a Nvidia RTX A6000. No early stopping was used.

[0176] The dataset of 3204 patients was conventionally split into training (N = 2213), validation (N = 505), and test cohorts (N = 486) - Table 2. One of our main objectives being to test the hypothesis of generalization of the trained network to unseen machines, data splitting was performed in a controlled manner. Thus, the data were stratified so as 18 different irradiation machines formed the training set (including classic linear accelerators and cobalt 60 devices), while 2 unseen machines were part of the validation set exclusively (one cobalt 60 unit, and one Sagittaire linear accelerator operating at 25 MV). The test set was divided into 5 subcohorts, including two classic linear accelerators (names of accelerator model unavailable) operating at 6 MV and 16 MV, two cobalt 60 units (Alcyon and Mobiletron), and finally a betatron operating at 1.25 MV, 10 MV, 12 MV, 14 MV, 16 MV, 18 MV, 22 MV and 32 MV. All patients treated with betatron devices were voluntarily kept into the test set, because of its very specific design compared with a conventional linear accelerator or a cobalt unit. No stratification on clinical data was applied.

[0177] Because the root mean square deviation (RMSD, Equation 2) is used as performance evaluation metrics by a lot of research teams developing analytical models for out-of-field dose estimation (

[0052] ;

[0078] ,

[0079] ,

[0080] ,

[0081] ) this measure was selected for the reporting of the results. To study the impact of the field size on the neural network performance, the results were also analyzed in subgroups, for which the threshold corresponded to the median size of the in-field volumes in the training set (3767 cm3). Similarly, RMSD values were computed in two different zones: we differentiated the area outside the radiation field into a near-field area and a far-field area. The 1% isodose was chosen arbitrarily to distinguish these two zones.

[0178] All manipulations on data have been implemented in Python 3.7. Table 2: Distribution of patients and devices included into training, validation and testing processes.

[0179] Results

[0180] Best performances were achieved at epoch 467 / 600. Table 3 presents the RMSD results obtained on training, validation and test sets for hyperparameters presented in the material and methods section.

[0181] Table 3: Performances of the 3D U-Net for the out-of-field dose map estimation task.

[0182] RMSD of 0.28 ± 0.08 and 0.41 ± 0.26 cGy.Gy1were obtained for the training and validation datasets, respectively. Values of 0.27 ± 0.06, 0.26 ± 0.07, 0.28 ± 0.06, 0.30 ± 0.12 and 0.65 ± 0.35 cGy.Gy1were achieved for the 6 MV linac, 16 MV linac,60Co-Alcyon,60Co-Mobiletron, and Betatron devices test sets, respectively, demonstrating overall performance similar to or better than that of the validation set, except for the fifth test set, corresponding to the Betatron device. The results observed in the validation and test sets as a function of distance from the irradiation field show better RMSD values from the field than close to it, except for the fourth (corresponding to the Mobiletron device) and fifth test sets. For example, far from the field RMSD values of 0.22 ± 0.12 and 0.35 ± 0.27 cGy.Gy1are reported for respectively the validation and the tests sets, while 0.49 ± 0.29 and 0.43 ± 0.22 cGy.Gy1are reported for area close from the field. Finally, the results of the validation set comparing dose maps for large fields and small fields suggest that the neural network performs better for large irradiation fields (0.37 ± 0.23 cGy.Gy1to be compared with 0.43 ± 0.28 cGy.Gy1), but this difference is less pronounced than the differences observed in the previous results between areas close to the field and areas far from the field, especially when taking into account standard deviation of RMSD results. This trend is also observed in the 5th test set (0.35 ± 0.14 cGy.Gy1to be compared with 0.66 ± 0.37 cGy.Gy1), while the other four test sets showing fairly similar results between large and small irradiation fields.

[0183] Histograms of the RMSD metric computed patient-wise on validation set and on test sets were plotted, leading to heavy-tailed distributions, and more specifically to log-normal distributions. On the basis of these figures, an RMSD threshold value of 0.6 cGy.Gy1was considered to separate good from poor out-of-field dose reconstructions. 87 out of 505 patients showed weaker performances in the validation set; 85 of whom being treated with a single device: a Sagittaire linear accelerator operating at 25 MV. This value was equal to 66 (out of 486 patients) in the test set, with 63 of the 66 patients identified having been treated with the Betatron accelerator.

[0184] Out-of-field dose maps were obtained for 3 representative patients: one with good performance (Patient 1, RMSD = 0.16 cGy.Gy1), one with median performance (Patient 2, RMSD = 0.29 cGy.Gy1) and one with poor performance (Patient 3, RMSD = 1.00 cGy.Gy1). Patient 1 (Test set 1 - RMSD= 0.16 cGy.Gy-1, into the 5% percentile of the distribution) corresponds to a 15-year-old girl diagnosed for a primary pathology of retinoblastoma in 1982 and treated with a Neptune 6 MV device. Patient 2 (Test set 3 - RMSD= 0.29 cGy.Gy-1, in ± 1% of median of the distribution) is also a 15-year-old girl who was diagnosed and treated in 1982 for an intracranial neoplasm using an Alcyon 60Co device. Patient 3 (Test set 5 - RMSD= 1.00 cGy.Gy-1, below 95% percentile of the distribution) is a 5-year-old girl diagnosed in 1982 treated the same year with a 1.25 MV Betatron device for an astrocytoma. For these 3 patients, applying our entire pre-processing pipeline including data loading, neural network application and data saving takes an average computation time of 11.1 seconds, 4.4 seconds being necessary for the out-of- field dose calculation itself.

[0185] For Patients 1 and 2, doses are well predicted in the near-field dose gradient, with mean relative dose differences, evaluated along the profiles, equal to 23.1 % and 26.3 % up to 20 cm from the edge of the field. For Patient 1, the continuous component far from the field is correctly reconstructed (average dose of 0.23 Gy for the ground truth compared with 0.26 Gy for the prediction by the network between 20 and 155 cm), but with a jump in dose that is not correctly predicted. For patient 2, the continuous component far from the field is not as well restored (average dose of 0.06 Gy for ground truth compared with 0.30 Gy for network prediction between 20 and 153 cm). For patient 3, treated on a Betatron machine, we note that the neural network has weaknesses in dose prediction not only in areas close to the field, but also in more distant areas, where, for example, it fails to predict the local increase in dose on the patient's legs associated with this specific linear accelerator.

[0186] Head-foot dose profiles of the whole-body dose maps predicted by the U-Net 3D network was compared with the dose profile of the associated ground truth dose map. Profiles were plotted only in the out-of-field dose areas, for three same patients. The doses were normalized to the maximum value obtained from the ground truth out-of-field dose maps (analytical model).

[0187] Finally, a link was established between the doses predicted by the network and the ground truth doses for each of the 3 patients, the objective being a linear curve passing through 0 with a slope of 1. The visualization confirms the observations made earlier from the profiles.

[0188] Discussion

[0189] Our aim was to prove the feasibility of out-of-field dose prediction for megavoltage photon irradiations using a deep learning neural network, while demonstrating that this approach is an appropriate response to the limitations in terms of computing time, difficulties in accessing experimental measurements and the lack of versatility inherent in analytical and MC methods, which ultimately limit access to out-of-field dose maps for routine clinical use.

[0190] Firstly, the results were analyzed in terms of RMSD, which is the most popular metric for assessing out-of-field doses. Conventionally, RMSD values were calculated in the training, validation and test sets, in order to assess the ability of the algorithm to generalize to the anatomy of new patients, new tumor locations, and new irradiation geometries, for new irradiation devices. The results are rather encouraging, with RMSD values of the same order of magnitude in the test sets (mean value of 0.39 ± 0.25 cGy.Gy-1 in the test set) as in the validation set (0.41 ± 0.26 cGy.Gy-1) or the training set (0.28 ± 0.08 cGy.Gy-1) (Table 3), suggesting that the neural network has acquired a strong degree of robustness and generalization. However, a closer look at the results for the different test sets shows that significantly lower results were observed for test set 5, which corresponds to the Betatron accelerator. In addition, most of the poorest results in the validation set concerned patients treated with the Sagittaire accelerator operating at 25 MV. As a reminder, the highest voltage of the linear accelerators considered in the training set was equal to 20 MV. These results logically highlight the fact that the generalization capabilities of the neural network cannot be extended to non-conventional or highly atypical linear accelerators, as long as they have not been considered during the training phase, i.e. with different shielding properties and internal geometries. Similarly, the ability to generalize is limited at very high voltages, as this implies in particular an increase in the pair production crosssection. For example, the Betatron device, which more closely resembles cyclotron systems than conventional devices, and the unique very high-voltage operation of the Sagittaire facility in our dataset contribute to this limited generalization capability. Naturally, a more diversified training database would make it possible to improve the generalization ability of the neural network.

[0191] Despite achieving better RMSD results in areas distant from the field compared to those near the field (Table 3), the neural network does not seem to correctly recover very low dose values. This consistent pattern observed in all dose maps generated by the neural network can be explained by two factors. Firstly, the chosen loss function (MSE) tends to minimize the absolute differences in dose between the predicted values and the ground truth. While some dose differences close to the field can be tolerated, because they are moderate in terms of relative differences, these same differences become more problematic when it comes to predicting very low doses far from the field. Relative MSE was implemented and investigated in order to bypass this limitation, but the results were inconclusive. Secondly, it is reasonable to assume that near-field doses, mainly influenced by the patient scatter component, are comparatively easier to predict for the network than the other components. In fact, this component depends mainly on the dose within the irradiation field and the irradiated volume, which is information readily available in the in-field dose map. On the other hand, very low doses in remote regions are mainly influenced by the leakage component, which is certainly only very partially present in the input data, apart from the fact that it is a signature of the accelerator. This observation also explains the association between larger irradiation fields and improved values (Table 3), thanks to more usable information available. Although some publications using analytical models for out-of-field dose estimation reported RMSD results similar to the performance achieved by our neural network (such as 0.91 cGy.Gy-1 and 1.67 cGy.Gy-1

[0081] , 0.75 cGy.Gy-1

[0082] , 4.1 cGy.Gy-1, 5.6 cGy.Gy-1, 4.6 cGy.Gy-1 and 6.5 cGy.Gy-1

[0078] , 1.04 cGy.Gy-1

[0079] , 3.7 mGy.Gy-1

[0083] , : 0.094 cGy.MU-1, 0.279 cGy.MU-1, and 0.410 cGy.MU-1

[0080] ), making a direct comparison is not relevant. Indeed, apart from the differences in the normalization process employed, which is a general problem in the context of out-of-field dose evaluation (

[0052] ), our analysis involves comparing the predictions of a DL neural network to out-of-field dose maps obtained from analytical computation, while the studies previously mentioned compared the predictions of analytical models with experimental measurements or MC simulation. The next stage of our work will therefore be to carry out experimental measurements in order to compare them with the predictions of the network.

[0192] The use of MC simulation with our modern hardware capabilities also now seems to be a good candidate for the generation of a learning database. Additionally, advancements in neural network strategies beyond our current architecture, such as adversarial auto-encoder, successfully used for extension of anatomopathological whole slide images

[0084] , hold potential for further refining our approach. It should be noted that various neural network architectures have been explored in the course of our experimentation, such as the 3D Unet transformer, and in the end it was the one presented here that demonstrated the best performance.

[0193] Based on this proof of concept, we have shown that deep learning is a relevant tool for addressing the limitations of analytical methods or MC simulation for out of field dose estimation. Thanks to its generalization capabilities and short inference times of just a few seconds, this tool should make it possible to move forward for routine clinical application and mass application in retrospective studies.

[0194] III. Application to patients with thoracic cancer

[0195] Introduction

[0196] In the era of immuno-oncology, concepts such as "lymphocyte sparing radiotherapy"

[0074] or "immunologically-fitted radiotherapy"

[0085] are driving us to rethink radiotherapy in order to take better account of the immune system, and lymphocytes in particular, in radiotherapy planning. In this context, several tools have been developed to estimate the dose received by blood or lymphocytes during treatment. Among them, the EDIC (or sometimes EDRIC) score, for "Effective Dose to (circulating) Immune Cells", was introduced in 2017 for lung irradiations

[0086] , based on an approach previously described by Yovino et al.

[0087] , Since, numerous studies have shown a correlation between EDIC and clinical outcomes or lymphopenia markers for patients with esophageal

[0011] ,

[0088] ,

[0089] ,

[0090] ,

[0091] ,

[0092] ,

[0093] ,

[0094] ,

[0095] ,

[0096] , lung

[0097] ,

[0098] ,

[0099] ,

[0100] ,

[0101] ,

[0102] ,

[0103] ,

[0104] ,

[0105] ,

[0106] ,

[0107] ,

[0108] ,

[0109] , breast

[0110] ,

[0111] ,

[0112] ,

[0113] , prostate

[0114] cancers, and mediastinal Hodgkin lymphoma

[0112] , The score was also proposed as a potential dose constraint to mitigate radiation-related lymphopenia

[0115] ,

[0197] However, to what extent the value provided by EDIC can really be interpreted as an "effective dose to (circulating) immune cells"? Mathematically, the score evaluates an average dose to the blood contained in lung, heart and total body. It then makes the assumption that the radiation dose to the circulating blood is a surrogate for the dose to the (circulating) immune cells

[0011] ,

[0089] ,

[0090] ,

[0098] , To discuss this hypothesis, we compared the results calculated with the EDIC score with those of our LymphoDose framework

[0116] , in the case of lung and esophageal irradiation. LymphoDose is a framework based on compartment models combined with Monte Carlo simulation. It can be used to calculate both the dose to peripheral blood or the dose received by recirculating lymphocytes, taking into account the recirculation and homing processes between lymphoid organs.

[0198] Material and methods

[0199] Three patients with thoracic tumors who had received at least 25 fractions of RT were selected. Their characteristics are presented in Table 4.

[0200] Table 4: Description of the 3 patients used in the analysis. M: Male, F: Female, SCC: squamous cell carcinoma, NSCLC: non-small cell lung carcinoma, EQD2: Equivalent Dose in 2Gy fractions, *computed witha / n = 10

[0201] Pl P2 P3

[0202] The EDIC definition used was: n y ITDV

[0203] 457 ' 61.8 ■ 10 3

[0204] Where n is the number of irradiation fractions (n > 25). MLD, MHD, and ITDV are the mean lung dose, mean heart dose and the integral total dose volume (calculated as the mean external contour dose multiplying with the volume of the external contour

[0110] ), respectively. B}% = 0.12, B2% = 0.08, B3% = 0.45 and B4% = 0.35 represent the percentage of blood volume within lung, heart, great vessels, and small vessels / capillaries in all other organs, respectively, / q = 0.85 is a dose effectiveness factor small vessels / capillary, and 61.8 ■ 103cm3is the average total body volume, assuming average weight and density of 63 kg and 1.02 g. cm~3. Details of the derivation of this equation are given in the Appendix A of

[0097] , To take account of the mean liver dose (MID) in the EDIC score, Xu et al.

[0011] 1 added a term B5% ■ kr■ ■ MID with B5% = 0.15 (also used in

[0095] ,

[0112] ). This score had also been calculated and will be referred to as EDIC_Xu in the following.

[0205] The results obtained with EDIC were compared with those of our LymphoDose famework, whose implementation details are described in

[0116] , The model is composed of 2 interconnected compartment models, M4and M2respectively. M4describes the slow recirculation of lymphocytes between blood and several lymphoid organs (spleen, subcutaneous lymph nodes, mesenteric lymph nodes and Peyer's Patches)

[0022] , mediated by specific biological processes such as L-selectin receptor and Sphingosine-l-Phosphate (SIP) pathways. M2describes the blood flow between 28 organs of the human body

[0025] , A stochastic process was used to simulate the lymphocytes' pathway between the compartments of these models, and to estimate the dose they receive when they pass through irradiated organs during beam-on time. M4has been adapted for thoracic irradiation. The subcutaneous lymph nodes compartment was divided into 5 sub-compartments: right nodes, left nodes, mammalian chain, mediastinal and other subcutaneous lymph nodes. From a study based on the visible human dataset having segmented more than 1,200 lymph nodes, we estimated that these sub-compartments accounted for 3.6%, 3.9%, 0.2% and 10.1% and 82.2% of all subcutaneous lymph nodes respectively. The M4transition probability matrix was adapted accordingly (see Table 5 below):

[0206] For each patient, lungs (grouping the left and right lungs), heart, liver, "aorta" and "vena cava" (grouping the superior and inferior vena cava) were contoured, in addition to spleen and major lymph nodes area in the irradiation field: right nodes (grouping LI, L2, L3 right axillary, sus-clavicular and interpectoral right lymph nodes), left nodes (grouping LI, L2, L3 left axillary, sus-clavicular and interpectoral left lymph nodes), mammalian chain and mediastinal nodes. Contouring was performed automatically by Annotate (TheraPanacea, Paris, France)

[0117] , with the exception of the mediastinal lymph nodes, aorta and vena cava, which were contoured by a senior physician. For all of the contoured structures i, a Dose Volume Histogram DVHi, the volume Vtand the mean dose Dtwere computed using dicompyler-core version 0.5.6

[0118] ,

[0207] The M2compartments aorta and large arteries and large veins do not represent a well-defined anatomical structure, and the associated contours "aorta" and "vena cava" merit careful study. The ICRP Publication 89

[0025] indicates that these compartments represent respectively 6% and 18% of the total blood volume at each time (male and female). Considering a total blood volume of 5.3L for men (resp. 3.9L for women), these 2 compartments contain at each time approximately 0.318L = 318cm3 (resp. 0.234L = 234cm3) and 0.954L = 954cm3 (resp. 0.702L = 702cm3) for men (resp. for women). The strategy adopted to define the anatomical contours "aorta" and "vena cava" associated with these compartments was as follows: only the parts of the aorta (and its daughter vessels) and the superior and inferior vena cava (and their daughter vessels) included in the 5% isodose were contoured. The volume of these structures, denoted Vseg aortaand Vseg venarespectively, were calculated. By neglecting the thickness of the vessel walls, these volumes can be considered to be made up entirely of blood. Thus, the volumes 318cm3— Vseg aortaand 954cm3— Vseg venafor men (resp. 234cm3— ^seg_aortaet702cm3— Vseg venafor women) are located out-of-field (i.e. outside the 5% isodose) and will receive a low dose, considered negligible here. The DVH associated with these contours has therefore been corrected to take account of these volumes having received a zero dose, and will be noted DVHaorta an(^ large arteries and DVHiarge veinsrespectively.

[0208] For each patient, the out-of-field dose map have been computed using the in-house deep learning neural network used to estimate the OOF dose (cf. Example II above for evaluating OOF doses generated by photon external beam radiotherapy). Note that for each structure taken into account (the lungs (grouping together the left and right lungs), the heart, the liver, the "aorta" and the "vena cava" (grouping together the superior and inferior vena cava) were delineated, in addition to the spleen and the main lymph nodes in the irradiation field: right lymph nodes (grouping LI, L2, L3 axillary, supraclavicular and right interpectoral nodes), left lymph nodes (grouping LI, L2, L3 axillary, supra-clavicular and left interpectoral nodes), mammalian chain and mediastinal nodes), the "in-field" or "out-of-field" character depends on the patient's irradiation plan. For example, for a patient (Rl) irradiated on the left lung, the left axillary nodes may be in-field, while the right axillary nodes may be out-of-field.

[0209] The EDIC and EDIC_Xu scores were calculated for each patient. Two scenarios were studied with LymphoDose. Scenario Hl considered that lymphocytes circulate only in peripheral blood, meaning that only the M2model is used. To each M2compartment irradiated in the field was assigned a DVH: DVHiung with bronchial and pulmonary, DVHheartwith right heart, left heart and coronary, DVHiiverwith liver, DVHa*orta and large arterieswith aorta and large arteries and DVHiarge veinswith large veins. Scenario H3 also considered that lymphocytes can transiently leave the peripheral blood and recirculate between different lymphoid organs, in addition to the dose received by out-of-field lymphoid structures. The Mi and M2models was interconnected, and each compartment of Mi irradiated in the field was also associated with a DVH: DVHspieenwith spleen, DVHright nodeswith right nodes, DV nodes with left nodes, ^f^f^mammalian chain with mammalian chain and DV ^mediastinaZ with mediastinal. Results were generated with 104and 105lymphocytes for Hl and H2 respectively.

[0210] Results

[0211] The DVH of each structure taken into account in LymphoDose for the 3 patients are presented in Figure 7. Because of a lower down located tumor, patient P3 had significantly higher liver and spleen doses than patients Pl and P2. On the contrary, patient P2 received a high dose at the left lymph nodes and right lymph nodes upper structures

[0212] Table 6: Different metrics summarizing the doses received by the blood (Hl) and the recirculating lymphocyte pool (H3). The mean dose, median dose and D2% are reported

[0213] Pl P2 P3

[0214] The DVH generated by LymphoDose (Hl and H3) are shown in Figure 8. The Table 5 summarizes the parameters extracted from these DVH, in particular the mean dose for Hl and H2. For the 3 patients, Figure 9 compares the results obtained by the EDIC and EDIC_Xu scores, and the mean LymphoDose Hl and H3 doses. EDIC, EDIC-Xu and LymphoDose Hl gave close results for all patients. On the contrary, the differences in mean lymphocyte pool dose calculated by LymphoDose H3 was higher, with 1.40Gy, 2.67Gy and 3.40 for patients Pl, P2 and P3 respectively.

[0215] Discussion

[0216] In this preliminary study, we compared EDIC scores with several LymphoDose framework-based metrics. These initial results are encouraging, as they suggest that LymphoDose could better discriminate between patients in terms of RIL, by taking lymphoid organs into account. The ongoing study on an extended cohort is intended to confirm this hypothesis and demonstrate the enhanced predictive capability of RIL for LymphoDose.

[0217] The EDIC score assumes that the radiation dose received by the peripheral blood is a surrogate for the dose received by the immune cells, and by lymphocytes in particular. However, this assumption is questionable for several reasons.

[0218] On one hand, if we consider only the resident lymphocytes in the organs considered, there is no reason why their distribution should be the same as that of the blood volume (the B% factors in the equation). For example, the lung contains between 3% to 4% of all immune cells in the human body

[0021] and about 2% of lymphocytes

[0022] , values very different from the 12% of blood volume used in Moreover, the EDIC score does not take into account the most significant immunological organs: the bone marrow, spleen and lymph nodes

[0021] , However, Zhang et al.

[0091] showed that the thoracic duct mean dose alone had a stronger predictive ability for lymphocyte count drop than the EDIC score for patients with lung cancer (areas under the curve = 0.72 (p < 0.001) v.s. 0.62 (p = 0.03)). To remedy this shortcoming, they proposed a modified EDIC (mEDIC) taking into account the mean thymus dose and the mean thoracic duct dose. Then, for patients who received esophageal irradiation, Liu et al

[0092] showed that EDIC was no longer significantly associated with overall survival (OS) when the V20Gy dose to the thoracic vertebrae was taken into account in multivariate Cox regression, but remained correlated with absolute lymphocyte count. Even if the relative contribution of the bone marrow dose to severe lymphopenia is debated

[0119] ,

[0120] ,

[0121] , it plays a role in lymphocyte recovery after treatment

[0094] , Then, Saito et al.

[0122] showed that the dose to spleen was an independent factors negatively influencing the absolute lymphocyte count at nadir for patients with oesophageal cancer. Finally, the negligence of lymph nodes is all the more problematic as they contain a huge numbers of lymphocytes (about 40% of the total pool

[0020] ), while some of them present in the irradiation field receiving doses that can be very high. Thus, as our results show, 2 patients (Pl and P2) who was relatively equivalent in terms of EDIC and dose to circulating blood (LymphoDose Hl) are no longer so when the recirculation and dose to lymphoid structures is taken into account (LymphoDose H2). Moreover, patient P3, who had the lowest EDIC (4.92 Gy), has the highest lymphocyte dose (...), partly due to the fact that he received an high average spleen dose of 8.4 Gy.

[0219] On the other hand, the EDIC score may also inadequately assess the dose to recirculating lymphocytes present in the blood. First, lymphocyte migration and retention in a specific tissue is not strictly determined by blood flow (the A% factors used to construct EDIC, see Appendix A of

[0097] ). This is particularly true for the lung and liver: a mathematical model based on rats data showed that 0.7% and 3.3%, of cardiac output goes to lung and liver respectively, which was in contrast to 78% and 17% of lymphocyte entry probability for these tissues

[0123] , Then, EDIC does not take into account lymphocyte recirculation and homing processes in lymphoid organs. The peripheral blood contains only between 2% to 5% of the total lymphocyte pool at any time

[0021] , which can be renewed up to 11 times a day as a result of these recirculation processes

[0068] , So, the lymphocyte present in the blood will potentially not be the same from one fraction to the next, and the dose does not simply accumulate at each fraction for a given circulating cell. This explains why, even adding the sometimes-high dose to lymph nodes, H2 doses are lower than Hl. Thus, as our results clearly show, while EDIC enables us to estimate an order of magnitude of the dose received by peripheral blood, it is not correct to extrapolate this dose to the lymphocytes it contains.

[0220] BIBLIOGRAPHIC REFERENCES

[0221] [1] E. Deutsch, C. Chargari, L. Galluzzi, et G. Kroemer, « Optimising efficacy and reducing toxicity of anticancer radioimmunotherapy », Lancet Oncol., vol. 20, n° 8, p. e452-e463, aout 2019, doi: 10.1016 / S1470-2045(19)30171-8.

[0222] [2] « Common Terminology Criteria for Adverse Events (CTCAE) », p. 147, 2017.

[0223] [3] S. Mallick et al., « A Systematic Review and Meta-analysis of the Impact of Radiation-Related Lymphopenia on Outcomes in High-Grade Gliomas », South Asian J. Cancer, p. s-0042-1753504, aout 2022, doi: 10.1055 / s-0042-1753504.

[0224] [4] Y. Zhang et al., « Prognostic Value and Risk Factors of Treatment-Related Lymphopenia in Malignant Glioma Patients Treated With Chemoradiotherapy: A Systematic Review and MetaAnalysis », Front. Neurol., vol. 12, p. 726561, janv. 2022, doi: 10.3389 / fneur.2021.726561.

[0225] [5] F. de Kermenguy et al., « Radio-induced lymphopenia in the era of anti-cancer immunotherapy », Int. Rev. Cell Mol. Biol., vol. 378, p. 1-30, 2023, doi: 10.1016 / bs.ircmb.2023.03.002.

[0226] [6] Y. El Houat, C. Massard, V. Quillien, R. de Crevoisier, et J. Castelli, « Meta-analysis and Critical Review: Association Between Radio-induced Lymphopenia and Overall Survival in Solid Cancers », Adv. Radiat. Oncol., vol. 8, n° 2, p. 101038, mars 2023, doi: 10.1016 / j.adro.2022.101038.

[0227] [7] D. Dai, Q. Tian, Y. Shui, J. Li, et Q. Wei, « The impact of radiation induced lymphopenia in the prognosis of head and neck cancer: A systematic review and meta-analysis », Radiother. Oncol.

[0228] J. Eur. Soc. Then Radiol. Oncol., p. S0167-8140(22)00007-X, janv. 2022, doi: 10.1016 / j.radonc.2022.01.003. [8] R. Upadhyay et al., « Risk and impact of radiation related lymphopenia in lung cancer: A systematic review and meta-analysis », Radiother. Oncol., vol. 157, p. 225-233, avr. 2021, doi: 10.1016 / j.radonc.2021.01.034.

[0229] [9] B. Venkatesulu et al., « Lymphocyte Sparing Normal Tissue Effects in the Clinic (LymphoTEC): A Systematic Review of Dose Constraint Considerations to Mitigate Radiation-Related Lymphopenia in the Era of Immunotherapy », Radiother. Oncol., nov. 2022, doi: 10.1016 / j.radonc.2022.10.019.

[0230]

[0010] L. Galluzzi, M. J. Aryankalayil, C. N. Coleman, et S. C. Formenti, « Emerging evidence for adapting radiotherapy to immunotherapy », Nat. Rev. Clin. Oncol., vol. 20, n° 8, Art. n° 8, aout 2023, doi: 10.1038 / s41571-023-00782-x.

[0231]

[0011] C. Xu et al., « The impact of the effective dose to immune cells on lymphopenia and survival of esophageal cancer after chemoradiotherapy », Radiother. Oncol. J. Eur. Soc. Then Radiol. Oncol., vol. 146, p. 180-186, mai 2020, doi: 10.1016 / j.radonc.2020.02.015.

[0232]

[0012] J. Shin et al., « HEDOS— a computational tool to assess radiation dose to circulating blood cells during external beam radiotherapy based on whole-body blood flow simulations », Phys. Med. Biol., vol. 66, n° 16, p. 164001, aout 2021, doi: 10.1088 / 1361-6560 / acl6ea.

[0233]

[0013] A. Hammi, H. Paganetti, et C. Grassberger, « 4D blood flow model for dose calculation to circulating blood and lymphocytes », Phys. Med. Biol., vol. 65, n° 5, p. 055008, mars 2020, doi: 10.1088 / 1361-6560 / ab6c41.

[0234]

[0014] S. Xing et al., « A dynamic blood flow model to compute absorbed dose to circulating blood and lymphocytes in liver external beam radiotherapy », Phys. Med. Biol., vol. 67, n° 4, p. 045010, fevr. 2022, doi: 10.1088 / 1361-6560 / ac4da4.

[0235]

[0015] H. Paganetti, « A review on lymphocyte radiosensitivity and its impact on radiotherapy », Front. Oncol., vol. 13, 2023, Consulte le: 31 aout 2023. [En ligne], Disponible sur: https: / / www.frontiersin.org / articles / 10.3389 / fonc.2023.1201500

[0236]

[0016] W. Sung et B. Cho, « Modeling of radiation effects to immune system: a review », J. Korean Phys. Soc., aout 2022, doi: 10.1007 / s40042-022-00574-z.

[0237]

[0017] L. van Hooren et al., « Agonistic CD40 therapy induces tertiary lymphoid structures but impairs responses to checkpoint blockade in glioma », Nat. Common., vol. 12, n° 1, Art. n° 1, juill. 2021, doi: 10.1038 / S41467-021-24347-7.

[0238]

[0018] L. Blanchard et J.-P. Girard, « High endothelial venules (HEVs) in immunity, inflammation and cancer », Angiogenesis, vol. 24, n° 4, p. 719-753, 2021, doi: 10.1007 / sl0456-021-09792-8.

[0239]

[0019] V. V. Ganusov et M. Tomura, « Experimental and Mathematical Approaches to Quantify Recirculation Kinetics of Lymphocytes », in Mathematical, Computational and Experimental T Cell Immunology, C. Molina-Pan's et G. Lythe, Ed., Cham: Springer International Publishing, 2021, p. 151-169. doi: 10.1007 / 978-3-030-57204-4_10.

[0240]

[0020] V. V. Ganusov et R. J. De Boer, « Do most lymphocytes in humans really reside in the gut? », Trends Immunol., vol. 28, n° 12, p. 514-518, dec. 2007, doi: 10.1016 / j.it.2007.08.009.

[0241]

[0021] R. Sender et al., « The total mass, number, and distribution of immune cells in the human body », Proc. Natl. Acad. Sci., vol. 120, n° 44, p. e2308511120, oct. 2023, doi: 10.1073 / pnas.2308511120.

[0242]

[0022] V. V. Ganusov et J. Auerbach, « Mathematical Modeling Reveals Kinetics of Lymphocyte Recirculation in the Whole Organism », PLOS Comput. Biol., vol. 10, n° 5, p. el003586, mai 2014, doi: 10.1371 / journal.pcbi.l003586.

[0243]

[0023] W. L. Ford, « Lymphocytes. 3. Distribution. Distribution of lymphocytes in health. », J. Clin. Pathol. Suppl. (R. Coll. Pathol.), vol. 13, p. 63-69, 1979.

[0244]

[0024] J.-Y. Jin et al., « A framework for modeling radiation induced lymphopenia in radiotherapy », Radiother. Oncol., vol. 144, p. 105-113, mars 2020, doi: 10.1016 / j.radonc.2019.11.014.

[0245]

[0025] ICRP Publication 89, « Basic Anatomical and Physiological Data for Use in Radiological Protection Reference Values », 2002, [En ligne], Disponible sur: https: / / www.icrp. org / publication.asp?id=icrp%20publication%2089

[0026] C. Beekman et al., « A stochastic model of blood flow to calculate blood dose during radiotherapy », Phys. Med. Biol., oct. 2023, doi: 10.1088 / 1361-6560 / ad02d6.

[0246]

[0027] C. Hall, E. Lueshen, A. Mosat', et A. A. Linninger, « Interspecies scaling in pharmacokinetics: a novel whole-body physiologically based modeling framework to discover drug biodistribution mechanisms in vivo », J. Pharm. Sci., vol. 101, n° 3, p. 1221-1241, mars 2012, doi: 10.1002 / jps.22811.

[0247]

[0028] S. S. Almberg, J. Frengen, et T. Lindmo, « Monte Carlo study of in-field and out-of-field dose distributions from a linear accelerator operating with and without a flattening-filter », Med. Phys., vol. 39, n° 8, p. 5194-5203, aout 2012, doi: 10.1118 / 1.4738963.

[0248]

[0029] R. M. Howell, S. B. Scarboro, S. F. Kry, et D. Z. Yaldo, « Accuracy of out-of-field dose calculations by a commercial treatment planning system », Phys. Med. Biol., vol. 55, n° 23, p. 6999-7008, dec. 2010, doi: 10.1088 / 0031-9155 / 55 / 23 / S03.

[0249]

[0030] R. M. Howell, S. B. Scarboro, P. J. Taddei, S. Krishnan, S. F. Kry, et W. D. Newhauser,

[0250] « Methodology for determining doses to in-field, out-of-field and partially in-field organs for late effects studies in photon radiotherapy », Phys. Med. Biol., vol. 55, n° 23, p. 7009-7023, dec. 2010, doi: 10.1088 / 0031-9155 / 55 / 23 / S04.

[0251]

[0031] B. Sanchez-Nieto, I. N. Lopez-Martinez, J. L. Rodriguez-Mongua, et I. Espinoza, « A simple analytical model for a fast 3D assessment of peripheral photon dose during coplanar isocentric photon radiotherapy », Front. Oncol., vol. 12, 2022, Consulte le: 12 octobre 2022. [En ligne], Disponible sur: https: / / www.frontiersin.org / articles / 10.3389 / fonc.2022.872752

[0252]

[0032] C. W. Schneider, W. D. Newhauser, L. J. Wilson, et R.-P. Kapsch, « A physics-based analytical model of absorbed dose from primary, leakage, and scattered photons from megavoltage radiotherapy with MLCs », Phys. Med. Biol., vol. 64, n° 18, p. 185017, sept. 2019, doi: 10.1088 / 1361-6560 / ab303a.

[0253]

[0033] L. Wang et G. X. Ding, « The accuracy of the out-of-field dose calculations using a model based algorithm in a commercial treatment planning system », Phys. Med. Biol., vol. 59, n° 13, p. N113-128, juill. 2014, doi: 10.1088 / 0031-9155 / 59 / 13 / N113.

[0254]

[0034] E. Goy et al., « The out-of-field dose in radiation therapy induces delayed tumorigenesis by senescence evasion », eLife, vol. 11, p. e67190, mars 2022, doi: 10.7554 / eLife.67190.

[0255]

[0035] M. Hauptmann et al., « Brain cancer after radiation exposure from CT examinations of children and young adults: results from the EPI-CT cohort study », Lancet Oncol., vol. 24, n° 1, p. 45-53, janv. 2023, doi: 10.1016 / 51470-2045(22)00655-6.

[0256]

[0036] J. Colnot, S. Zefkili, R. Gschwind, et C. Huet, « Out-of-field doses from radiotherapy using photon beams: A comparative study for a pediatric renal treatment », J. Appl. Clin. Med. Phys., fevr. 2021, doi: 10.1002 / acm2.13182.

[0257]

[0037] S. Mutic et D. A. Low, « Whole-body dose from tomotherapy delivery », Int. J. Radiat. Oncol. Biol. Phys., vol. 42, n° 1, p. 229-232, aout 1998, doi: 10.1016 / s0360-3016(98)00199-0.

[0258]

[0038] Y. Tao et al., « Comparison of dose contribution to normal pelvic tissues among conventional, conformal and intensity-modulated radiotherapy techniques in prostate cancer », Acta Oncol. Stockh. Swed., vol. 47, n° 3, p. 442-450, 2008, doi: 10.1080 / 02841860701666055.

[0259]

[0039] J. Yoon, D. Heins, X. Zhao, M. Sanders, et R. Zhang, « Measurement and modeling of out-of-field doses from various advanced post-mastectomy radiotherapy techniques », Phys. Med. Biol., vol. 62, n° 23, p. 9039-9053, nov. 2017, doi: 10.1088 / 1361-6560 / aa94b5.

[0260]

[0040] M. L. De Bruin et al., « Breast Cancer Risk in Female Survivors of Hodgkin's Lymphoma: Lower Risk After Smaller Radiation Volumes », J. Clin. Oncol., vol. 27 , n° 26, p. 4239-4246, sept. 2009, doi: 10.1200 / JC0.2008.19.9174.

[0261]

[0041] E. J. Hall, « Intensity-modulated radiation therapy, protons, and the risk of second cancers », Int. J. Radiat. Oncol. Biol. Phys., vol. 65, n° 1, p. 1-7, mai 2006, doi: 10.1016 / j.ijrobp.2006.01.027.

[0262]

[0042] E. J. Hall et C.-S. Wuu, « Radiation-induced second cancers: the impact of 3D-CRT and IMRT », Int. J. Radiat. Oncol. Biol. Phys., vol. 56, n° 1, p. 83-88, mai 2003, doi: 10.1016 / S0360- 3016(03)00073-7.

[0043] M.-C. JahreiR et al., « The Risk of Second Primary Cancers in Prostate Cancer Survivors Treated in the Modern Radiotherapy Era », Front. Oncol., vol. 10, 2020, Consulte le: 31 mai 2023. [En ligne], Disponible sur: https: / / www.frontiersin.org / articles / 10.3389 / fonc.2020.605119

[0263]

[0044] M. Xiang, D. T. Chang, et E. L. Pollom, « Second cancer risk after primary cancer treatment with three-dimensional conformal, intensity-modulated, or proton beam radiation therapy », Cancer, vol. 126, n° 15, p. 3560-3568, aout 2020, doi: 10.1002 / cncr.32938.

[0264]

[0045] P. Lambin et al., « Lymphocyte-Sparing Radiotherapy: The Rationale for Protecting Lymphocyterich Organs When Combining Radiotherapy With Immunotherapy », Semin. Radiat. Oncol., vol. 30, n° 2, p. 187-193, avr. 2020, doi: 10.1016 / j.semradonc.2019.12.003.

[0265]

[0046] S. Cai et al., « Impact of Radiation Dose to Circulating Immune Cells on Tumor Control and Survival in Esophageal Cancer », Cancer Blather. Radiopharm., dec. 2021, doi: 10.1089 / cbr.2021.0250.

[0266]

[0047] D. Heylmann, V. Ponath, T. Kindler, et B. Kaina, « Comparison of DNA repair and radiosensitivity of different blood cell populations », Sci. Rep., vol. 11, n° 1, p. 2478, janv. 2021, doi: 10.1038 / S41598-021-81058-1.

[0267]

[0048] N. Nakamura, Y. Kusunoki, et M. Akiyama, « Radiosensitivity of CD4 or CD8 positive human T- lymphocytes by an in vitro colony formation assay », Radiat. Res., vol. 123, n° 2, p. 224-227, aout 1990.

[0268]

[0049] J. Colnot et al., « A new Monte Carlo model of a Cyberknife® system for the precise determination of out-of-field doses », Phys. Med. Biol., vol. 64, n° 19, p. 195008, oct. 2019, doi: 10.1088 / 1361-6560 / ab38e5.

[0269]

[0050] B. Sanchez-Nieto, K. N. Medina-Ascanio, J. L. Rodn'guez-Mongua, E. Doerner, et I. Espinoza,

[0270] « Study of out-of-field dose in photon radiotherapy: A commercial treatment planning system versus measurements and Monte Carlo simulations », Med. Phys., juin 2020, doi: 10.1002 / mp.14356.

[0271]

[0051] U. Schneider et al., « Accuracy of out-of-field dose calculation of tomotherapy and cyberknife treatment planning systems: a dosimetric study », Z. Med. Phys., vol. 24, n° 3, p. 211-215, sept. 2014, doi: 10.1016 / j.zemedi.2013.10.008.

[0272]

[0052] N. Benzazon et al., « Analytical models for external photon beam radiotherapy out-of-field dose calculation: a scoping review », Front. Oncol., vol. 13, 2023, Consulte le: 29 mai 2023. [En ligne], Disponible sur: https: / / www.frontiersin.org / articles / 10.3389 / fonc.2023.1197079

[0273]

[0053] F. Shi et al., « Deep learning empowered volume delineation of whole-body organs-at-risk for accelerated radiotherapy », Nat. Common., vol. 13, n° 1, Art. n° 1, nov. 2022, doi:

[0274] 10.1038 / s41467-022-34257-x.

[0275]

[0054] C. McIntosh et al., « Clinical integration of machine learning for curative-intent radiation treatment of patients with prostate cancer », Nat. Med., vol. 27, n° 6, p. 999-1005, juin 2021, doi: 10.1038 / s41591-021-01359-w.

[0276]

[0055] M. Lerner, J. Medin, C. Jamtheim Gustafsson, S. Alkner, et L. E. Olsson, « Prospective Clinical Feasibility Study for MRI-Only Brain Radiotherapy », Front. Oncol., vol. 11, p. 812643, 2021, doi: 10.3389 / fonc.2021.812643.

[0277]

[0056] E. Huynh et al., « Artificial intelligence in radiation oncology », Nat. Rev. Clin. Oncol., vol. 17, n° 12, Art. n° 12, dec. 2020, doi: 10.1038 / s41571-020-0417-8.

[0278]

[0057] K. R. Kase, G. K. Svensson, A. B. Wolbarst, et M. A. Marks, « Measurements of dose from secondary radiation outside a treatment field », Int. J. Radiat. Oncol. Biol. Phys., vol. 9, n° 8, p. 1177-1183, aout 1983, doi: 10.1016 / 0360-3016(83)90177-3.

[0279]

[0058] M. A. Benadjaoud et al., « A multi-plane source model for out-of-field head scatter dose calculations in external beam photon therapy », Phys. Med. Biol., vol. 57, n° 22, p. 7725-7739, nov. 2012, doi: 10.1088 / 0031-9155 / 57 / 22 / 7725.

[0280]

[0059] N. Bougas et al., « Surveillance after childhood cancer: are survivors with an increased risk for cardiomyopathy regularly followed-up? », Br. J. Cancer, vol. 129, n° 8, Art. n° 8, oct. 2023, doi: 10.1038 / S41416-023-02400-0.

[0060] E. Qinlar, « Markov Renewal Theory », Adv. Appl. Probab., vol. 1, n° 2, p. 123-187, 1969, doi: 10.2307 / 1426216.

[0281]

[0061] « Julia: A Fresh Approach to Numerical Computing | SIAM Review ». Consulte le: 29 mai 2023. [En ligne] . Disponible sur: https: / / epubs-siam-org.ezproxy.universite-paris- saclay.fr / doi / 10.1137 / 141000671

[0282]

[0062] V. Gregoire et al., « Delineation of the neck node levels for head and neck tumors: a 2013 update. DAHANCA, EORTC, HKNPCSG, NCIC CTG, NCRI, RTOG, TROG consensus guidelines », Radiother. Oncol. J. Ear. Soc. Then Radiol. Oncol., vol. 110, n° 1, p. 172-181, janv. 2014, doi: 10.1016 / j.radonc.2013.10.010.

[0283]

[0063] N. Benzazon et al., « Analytical models for external photon beam radiotherapy out-of-field dose calculation: a scoping review », Front. Oncol., vol. 13, p. 1197079, 2023, doi:

[0284] 10.3389 / fonc.2023.1197079.

[0285]

[0064] R. M. Howell, S. B. Scarboro, S. F. Kry, et D. Z. Yaldo, « Accuracy of out-of-field dose calculations by a commercial treatment planning system », Phys. Med. Biol., vol. 55, n° 23, p. 6999-7008, dec. 2010, doi: 10.1088 / 0031-9155 / 55 / 23 / S03.

[0286]

[0065] S. M. Qatarneh, L-C. Kiricuta, A. Brahme, U. Tiede, et B. K. Lind, « Three-dimensional atlas of lymph node topography based on the visible human data set », Anat. Rec. B. NewAnat., vol. 289, n° 3, p. 98-111, mai 2006, doi: 10.1002 / ar.b.20102.

[0287]

[0066] K. R. Kase, G. K. Svensson, A. B. Wolbarst, et M. A. Marks, « Measurements of dose from secondary radiation outside a treatment field », Int. J. Radiat. Oncol. Biol. Phys., vol. 9, n° 8, p. 1177-1183, aout 1983, doi: 10.1016 / 0360-3016(83)90177-3.

[0288]

[0067] J. L. GOWANS, « The effect of the continuous re-infusion of lymph and lymphocytes on the output of lymphocytes from the thoracic duct of unanaesthetized rats. », Br. J. Exp. Pathol., vol. 38, n° 1, p. 67-78, 1957.

[0289]

[0068] M. J. Gough et M. R. Crittenden, « The paradox of radiation and T cells in tumors », Neoplasia N. Y. N, vol. 31, p. 100808, juin 2022, doi: 10.1016 / j.neo.2022.100808.

[0290]

[0069] A. Arina et al., « Tumor-reprogrammed resident T cells resist radiation to control tumors », Nat. Common., vol. 10, n° 1, p. 3959, sept. 2019, doi: 10.1038 / s41467-019-11906-2.

[0291]

[0070] O. A. Trowell, « The sensitivity of lymphocytes to ionising radiation », J. Pathol. Bacterial., vol. 64, n° 4, p. 687-704, oct. 1952, doi: 10.1002 / path.1700640403.

[0292]

[0071] D. Heylmann, V. Ponath, T. Kindler, et B. Kaina, « Comparison of DNA repair and radiosensitivity of different blood cell populations », Sci. Rep., vol. 11, n° 1, p. 2478, janv. 2021, doi: 10.1038 / S41598-021-81058-1.

[0293]

[0072] F. B. Geara et al., « Intrinsic radiosensitivity of normal human fibroblasts and lymphocytes after high- and low-dose-rate irradiation », Cancer Res., vol. 52, n° 22, p. 6348-6352, nov. 1992.

[0294]

[0073] C. Veres et al., « Retrospective reconstructions of active bone marrow dose-volume histograms », Int. J. Radiat. Oncol. Biol. Phys., vol. 90, p. 1216-1224, 2014, doi: 10.1016 / j.ijrobp.2014.08.335.

[0295]

[0074] P. Francois, C. Beurtheret, et A. Dutreix, « Calculation of the dose delivered to organs outside the radiation beams », Med. Phys., vol. 15, p. 879-883, 1988, doi: 10.1118 / 1.596170.

[0296]

[0075] O. Qigek, A. Abdulkadir, S. S. Lienkamp, T. Brox, et O. Ronneberger, « 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation », ArXivl60606650 Cs, juin 2016, Consulte le: 29 juillet 2021. [En ligne], Disponible sur: http: / / arxiv.org / abs / 1606.06650

[0297]

[0076] T. Falk, D. Mai, R. Bensch, et al., « U-Net: deep learning for cell counting, detection, and morphometry », Nat. Methods, vol. 16, p. 67-70, 2019, doi: 10.1038 / s41592-018-0261-2.

[0298]

[0077] O. Ronneberger, P. Fischer, et T. Brox, « U-net: Convolutional networks for biomedical image segmentation ». 2015.

[0299]

[0078] M. De Saint-Hubert et al., « Out-of-field doses from pediatric craniospinal irradiations using 3D- CRT, IMRT, helical tomotherapy and electron-based therapy », Phys. Med. Biol., vol. 62, p. 5293-5311, 2017, doi: 10.1088 / 1361-6560 / aa6c9e.

[0079] L. J. Jagetic et W. D. Newhauser, « A simple and fast physics-based analytical method to calculate therapeutic and stray doses from external beam, megavoltage x-ray therapy », Phys. Med. Biol., vol. 60, p. 4753-4775, 2015, doi: 10.1088 / 0031-9155 / 60 / 12 / 4753.

[0300]

[0080] B. Sanchez et al., « Analytical model for photon peripheral dose estimation in radiotherapy treatments », Biomed. Phys. Eng. Express, vol. 1, p. 045205, 2015.

[0301]

[0081] P. J. Taddei et al., « Analytical model for out-of-field dose in photon craniospinal irradiation », Phys. Med. Biol., vol. 58, p. 7463-7479, 2013, doi: 10.1088 / 0031-9155 / 58 / 21 / 7463.

[0302]

[0082] K. J. Gallagher et al., « Supplemental computational phantoms to estimate out-of-field absorbed dose in photon radiotherapy », Phys. Med. Biol., vol. 63, p. 025021, 2018, doi: 10.1088 / 1361- 6560 / aa9838.

[0303]

[0083] L. J. Wilson et al., « Method to quickly and accurately calculate absorbed dose from therapeutic and stray photon exposures throughout the entire body in individual patients », Med. Phys., vol. 47, p. 2254-2266, 2020, doi: 10.1002 / mp.14018.

[0304]

[0084] J. Boyd, M. Liashuha, E. Deutsch, N. Paragios, S. Christodoulidis, et M. Vakalopoulou, « Selfsupervised representation learning using visual field expansion on digital pathology », in 2021 IEEE / CVF international conference on computer vision workshops (ICCVW), IEEE Computer Society, 2021, p. 639-647. doi: 10.1109 / ICCVW54120.2021.00077.

[0305]

[0085] D. Morel, C. Robert, N. Paragios, V. Gregoire, et E. Deutsch, « Translational frontiers and clinical opportunities of immunologically-fitted radiotherapy », Clin. Cancer Res., mars 2024, doi:

[0306] 10.1158 / 1078-0432. CCR-23-3632.

[0307]

[0086] J. Y. Jin et al., « Higher Radiation Dose to Immune System is Correlated With Poorer Survival in Patients With Stage III Non-small Cell Lung Cancer: A Secondary Study of a Phase 3 Cooperative Group Trial (NRG Oncology RTOG 0617) », Int. J. Radiat. Oncol. Biol. Phys., vol. 99, n° 2, p. S151-S152, oct. 2017, doi: 10.1016 / j.ijrobp.2017.06.351.

[0308]

[0087] S. Yovino, L. Kleinberg, S. A. Grossman, M. Narayanan, et E. Ford, « The Etiology of Treatment- related Lymphopenia in Patients with Malignant Gliomas: Modeling Radiation Dose to Circulating Lymphocytes Explains Clinical Observations and Suggests Methods of Modifying the Impact of Radiation on Immune Cells », Cancer Invest., vol. 31, n° 2, p. 140-144, fevr. 2013, doi: 10.3109 / 07357907.2012.762780.

[0309]

[0088] S. Cai et al., « Impact of Radiation Dose to Circulating Immune Cells on Tumor Control and Survival in Esophageal Cancer », Cancer Blather. Radiopharm., dec. 2021, doi: 10.1089 / cbr.2021.0250.

[0310]

[0089] T. H. So et al., « Lymphopenia and Radiation Dose to Circulating Lymphocytes With Neoadjuvant Chemoradiation in Esophageal Squamous Cell Carcinoma », Adv. Radiat. Oncol., vol. 5, n° 5, p. 880-888, 2020, doi: 10.1016 / j.adro.2020.03.021.

[0311]

[0090] J. Qiu et al., « Higher radiation dose on immune cells is associated with radiation-induced lymphopenia and worse prognosis in patients with locally advanced esophageal squamous cell carcinoma », Front. Immunol., vol. 14, p. 1066255, mai 2023, doi:

[0312] 10.3389 / fimmu.2023.1066255.

[0313]

[0091] J. Zhang et al., « Dosimetric Effect of Thymus and Thoracic Duct on Radiation-Induced Lymphopenia in Patients With Primary Lung Cancer Who Received Thoracic Radiation », Adv. Radiat. Oncol., vol. 8, n° 6, p. 101260, 2023, doi: 10.1016 / j.adro.2023.101260.

[0314]

[0092] M. Liu, X. Li, H. Cheng, Y. Wang, et Y. Tian, « The Impact of Lymphopenia and Dosimetric Parameters on Overall Survival of Esophageal Cancer Patients Treated with Definitive Radiotherapy », Cancer Manag. Res., vol. 13, p. 2917-2924, mars 2021, doi: 10.2147 / CMAR.S297010.

[0315]

[0093] I. Tseng et al., « Less Irradiation to Lymphocyte-Related Organs Reduced the Risk of G4 Lymphopenia in Esophageal Cancer: Re-Analysis of Prospective Trials », The Oncologist, p. oyadl09, juin 2023, doi: 10.1093 / oncolo / oyadl09.

[0316]

[0094] I. Tseng et al., « Lymphocyte recovery from radiation-induced lymphopenia in locally advanced esophageal squamous cell carcinoma: correlations with prognosis and lymphocyte-related organs », Radiat. Oncol. Lond. Engl., vol. 18, n° 1, p. 172, oct. 2023, doi: 10.1186 / sl3014-023- 02354-w.

[0317]

[0095] M. Zhang, Z. Li, et Y. Yin, « Association of effective dose to immune cells and vertebral marrow dose with hematologic toxicity during neoadjuvant chemoradiotherapy in esophageal squamous cell carcinoma », BMC Cancer, vol. 24, n° 1, p. 779, juin 2024, doi: 10.1186 / sl2885- 024-12531-z.

[0318]

[0096] D. Luo et al., « The predictors of lymphopenia and its effects on survival in locally advanced esophageal squamous cell carcinoma », Cancer Biol. Then, vol. 25, n° 1, p. 2371632, dec. 2024, doi: 10.1080 / 15384047.2024.2371632.

[0319]

[0097] J.-Y. Jin et al., « Higher Radiation Dose to the Immune Cells Correlates with Worse Tumor Control and Overall Survival in Patients with Stage III NSCLC: A Secondary Analysis of RTOG0617 », Cancers, vol. 13, n° 24, p. 6193, dec. 2021, doi: 10.3390 / cancersl3246193.

[0320]

[0098] C. J. Ladbury, C. G. Rusthoven, D. R. Camidge, B. D. Kavanagh, et S. K. Nath, « Impact of Radiation Dose to the Host Immune System on Tumor Control and Survival for Stage III NonSmall Cell Lung Cancer Treated with Definitive Radiation Therapy », Int. J. Radiat. Oncol. Biol. Phys., vol. 105, n° 2, p. 346-355, oct. 2019, doi: 10.1016 / j.ijrobp.2019.05.064.

[0321]

[0099] N. S. McCall et al., « Impact of Radiation Dose to the Immune Cells in Unresectable or Stage III Non-Small Cell Lung Cancer in the Durvalumab Era », Radiother. Oncol., juill. 2022, doi: 10.1016 / j.radonc.2022.07.015.

[0322]

[0100] Y. Yu et al., « Impact of effective dose to immune cells (EDIC) on lymphocyte nadir and survival in limited-stage SCLC », Radiother. Oncol. J. Ear. Soc. Then Radiol. Oncol., vol. 162, p. 26-33, sept. 2021, doi: 10.1016 / j.radonc.2021.06.020.

[0323]

[0101] M. Thor et al., « Pre-treatment immune status predicts disease control in NSCLCs treated with chemoradiation and durvalumab », Radiother. Oncol. J. Eur. Soc. Then Radiol. Oncol., vol. 167, p. 158-164, dec. 2021, doi: 10.1016 / j.radonc.2021.12.016.

[0324]

[0102] C. Friedes et al., « The effective radiation dose to immune cells predicts lymphopenia and inferior cancer control in locally advanced NSCLC », Radiother. Oncol. J. Eur. Soc. Then Radiol. Oncol., vol. 190, p. 110030, janv. 2024, doi: 10.1016 / j.radonc.2023.110030.

[0325]

[0103] G. Yang et al., « Risk of on-treatment lymphopenia is associated with treatment outcome and efficacy of consolidation immunotherapy in patients with non-small cell lung cancer treated with concurrent chemoradiotherapy », Radiother. Oncol. J. Eur. Soc. Then Radiol. Oncol., vol.

[0326] 189, p. 109934, dec. 2023, doi: 10.1016 / j.radonc.2023.109934.

[0327]

[0104] C. Pasquier et al., « Impact of radiation on host immune system in patients treated with chemoradiotherapy and durvalumab consolidation for unresectable locally advanced non-small cell lung cancer », Front. Oncol., vol. 13, p. 1186479, 2023, doi: 10.3389 / fonc.2023.1186479.

[0328]

[0105] Z. Nowicka et al., « Dose-volume metric-based prediction of radiotherapy-induced lymphocyte loss in patients with non-small-cell lung cancer treated with modern radiotherapy techniques », Phys. Imaging Radiat. Oncol., p. 100593, mai 2024, doi: 10.1016 / j.phro.2024.100593.

[0329]

[0106] t. Kuncman et al., « Early Lymphocyte Levels and Low Doses Radiation Exposure of Lung Predict Lymphopenia in Radiotherapy for Lung Cancer », Front. Immunol., vol. 15, juill. 2024, doi: 10.3389 / fimmu.2024.1426635.

[0330]

[0107] F. Dionisi et al., « Dosimetric and NTCP advantages of robust proton therapy over robust VMAT for Stage III NSCLC in the immunotherapy era », Phys. Medico PM Int. J. Devoted Appl. Phys. Med. Biol. Off. J. Ital. Assoc. Biomed. Phys. AIFB, vol. 123, p. 103410, juill. 2024, doi: 10.1016 / j.ejmp.2024.103410.

[0331]

[0108] J. S. Patel et al., « Immune System Dose With Proton Versus Photon Radiotherapy for Treatment of Locally Advanced NSCLC », Int. J. Part. Then, vol. 12, p. 100016, juin 2024, doi: 10.1016 / j.ijpt.2024.100016.

[0332]

[0109] S. Monti, T. Xu, Z. Liao, R. Mohan, L. Celia, et G. Palma, « On the interplay between dosiomics and genomics in radiation-induced lymphopenia of lung cancer patients », Radiother. Oncol. J. Eur. Soc. Then Radiol. Oncol., vol. 167, p. 219-225, fevr. 2022, doi: 10.1016 / j.radonc.2021.12.038.

[0110] F. Chen et al., « Radiation Induced Lymphopenia Is Associated With the Effective Dose to the Circulating Immune Cells in Breast Cancer », Front. Oncol., vol. 12, p. 768956, 2022, doi:

[0333] 10.3389 / fonc.2022.768956.

[0334]

[0111] F. Chen et al., « Radiation Induced Lymphopenia is Associated with the Effective Dose to the Circulating Immune Cells (EDIC) for Breast Cancer », Int. J. Radiat. Oncol. Biol. Phys., vol. 108, n° 3, p. e57-e58, nov. 2020, doi: 10.1016 / j.ijrobp.2020.07.1115.

[0335]

[0112] P. Loap et al., « Proton therapy reduces the effective dose to immune cells in mediastinal Hodgkin lymphoma patients », Int. J. Part. Then, p. 100110, juin 2024, doi: 10.1016 / j.ijpt.2024.100110.

[0336]

[0113] P. Loap, J. Vu Bezin, L. De Marzi, et Y. Kirova, « Determinants of radiation dose to immune cells during breast radiotherapy », Strahlenther. Onkol. Organ Dtsch. Rontgengesellschaft Al, mai 2024, doi: 10.1007 / s00066-024-02240-8.

[0337]

[0114] M. Bare et al., « Lymphocyte-sparing pelvic radiotherapy for prostate cancer: An in-silico study », Phys. Imaging Radiat. Oncol., vol. 23, p. 127-133, juill. 2022, doi: 10.1016 / j.phro.2022.07.006.

[0338]

[0115] B. Venkatesulu et al., « Lymphocyte sparing normal tissue effects in the clinic (LymphoTEC): A systematic review of dose constraint considerations to mitigate radiation-related lymphopenia in the era of immunotherapy », Radiother. Oncol., vol. 177, p. 81-94, dec. 2022, doi: 10.1016 / j.radonc.2022.10.019.

[0339]

[0116] F. de Kermenguy et al., « LymphoDose: a lymphocyte dose estimation framework - application to brain radiotherapy », Phys. Med. Biol., 2024, doi: 10.1088 / 1361-6560 / ad3c8d.

[0340]

[0117] S. Stathakis et al., « Evaluation of Al vs. Clinical Experts SBRT-Thorax Computed Tomography OARs Delineation », Int. J. Radiat. Oncol. Biol. Phys., vol. 114, n° 3, p. el02-el03, nov. 2022, doi: 10.1016 / j.ijrobp.2022.07.897.

[0341]

[0118] A. Panchai et al., dicompyler / dicompyler-core: vO.5.6. (8 mai 2023). Zenodo. doi: 10.5281 / zenodo.7908530.

[0342]

[0119] « RE: Venkatesulu et al. (Letter to the Editor) - Cl inical Key ». Consulte le: 14 mars 2024. [En ligne] . Disponible sur: https: / / www.clinicalkey.fr / #! / content / playContent / l-s2.0- S0167814023000282?returnurl=https:%2F%2Flinkinghub.elsevier.com%2Fretrieve%2Fpii%2FS0 167814023000282%3Fshowall%3Dtrue&referrer=https:%2F%2Fpubmed.ncbi.nlm.nih.gov%2F

[0343]

[0120] « With regards to Ellsworth et al. on "Lymphocyte sparing normal tissue effects in the clinic LymphoTEC): A systematic review of dose constraint considerations to mitigate radiation-related lymphopenia in the era of immunotherapy" - Clinical Key ». Consulte le: 14 mars 2024. [En ligne], Disponible sur: https: / / www.clinicalkey.fr / #! / content / journal / l-s2.0- S0167814023000348#reflnSitub0005

[0344]

[0121] J. Ma et al., « The Role of Thoracic Vertebral Body Dosimetry in Minimizing Acute Hematologic Toxicities of Patients With Non-Small Cell Lung Cancer Receiving Lung Radiation Therapy and Immunotherapy », Int. J. Radiat. Oncol. Biol. Phys., vol. 119, n° 1, p. 78-89, mai 2024, doi: 10.1016 / j.ijrobp.2023.11.037.

[0345]

[0122] T. Saito et al., « Spleen Dose-Volume Parameters as a Predictor of Treatment-related Lymphopenia During Definitive Chemoradiotherapy for Esophageal Cancer », In Vivo, vol. 32, n° 6, p. 1519-1525, nov. 2018, doi: 10.21873 / invivo.ll409.

[0346]

[0123] V. V. Ganusov et M. Tomura, « Experimental and mathematical approaches to quantify recirculation kinetics of lymphocytes », 20 aout 2018, bioRxiv. doi: 10.1101 / 268326.

Claims

CLAIMS1. A method for determining the irradiation dose that has been or will be received by a lymphocyte L during a radiotherapy treatment performed on a target organ of a patient, said method comprising the steps of: a) collecting the following data associated to said radiotherapy treatment:- the radiation treatment planning applied to or scheduled for said patient,- the radiation reference doses applied to or scheduled for said patient, and- the number, location and volume of the lymphoid organs that have been or will be irradiated (if any), b) simulating the lymphocyte path between the several organs of the body, c) determining the irradiation dose received by said lymphocyte L during said radiotherapy treatment would it be circulating in peripheral blood, or present in or circulating between lymphoid organs, d) determining the irradiation dose received by at least one out-of-field lymphoid organ during said radiotherapy treatment, e) combining these data in order to determine the irradiation dose that has been or will be received by the lymphocytes of said patient during said radiotherapy treatment.

2. The method of claim 1, wherein step b) is performed by using a Markovian or a semi-Markovian calculation.

3. The method of any one of claims 1-2, wherein step d) is performed on at least one, preferably on all, out-of-field lymphoid organ chosen from spleen, subcutaneous lymph nodes, mesenteric lymph nodes, and Peyer's patches via analytical, Monte-Carlo or artificial intelligence-based methods.

4. The method of any one of claims 1-3, wherein step d) is performed by directly or indirectly entering data collected in step a) into a previously trained out of field (OOF) dose prediction model, which returns the irradiation dose received by at least one out-of-field lymphoid organ during said radiotherapy treatment, wherein the OOF dose prediction model is an artificial intelligence-based OOF dose prediction model, preferably a deep-learning OOF dose prediction model and more preferably aneural network-based OOF dose prediction model, which has been trained on at least 500 whole body dose maps from cancer patients treated by radiotherapy using different radiotherapy machines.

5. The method of claim 4, wherein:• the radiation treatment planning applied to or scheduled for said patient (Rl) comprises partial body medical imaging,• the radiation reference doses applied to or scheduled for said patient (R2) are provided as an in-field dose map from the treatment-planning system,• the method further comprises before step d) a step dO) of generating a whole body dose map from: o the patient's external contour, and o the in-field dose map from the treatment-planning system, and• Step d) determines the irradiation dose received by at least one out-of-field lymphoid organ during said radiotherapy treatment based on the whole body dose map generated in step dO) and a whole body patient's anatomy.

6. The method of any one of claims 1-5, wherein it is implemented by a computer.

7. The method of any one of claims 1-6, wherein step e) enables to calculate the percentage of lymphocytes that have been or that will be lethal ly irradiated in said patient during said radiotherapy treatment.

8. The method of any one of claims 1-6, wherein step e) enables to determine a dose volume histogram for the lymphocyte pool.

9. The method of any one of claims 1-8, wherein said target organ is the brain and said out-of-field lymphoid organs are the subcutaneous lymph nodes of the head-and-neck region.

10. The method of any one of claims 1-8, wherein:• The target organ is right lung, radiation treatment planning (RTP) is such that in-field thoracic lymphoid organs comprise bone marrow, right lymph nodes (including LI, L2, L3 right axillary, sus-clavicular and interpectoral right lymph nodes), mammalian chain lymph nodes and mediastinal lymph nodes and out-of-field thoracic lymphoid organs comprise spleen, left lymph nodes (including LI, L2, L3 left axillary, sus-clavicular and interpectoral left lymph nodes) and mesenteric lymph nodes, and the irradiation doses received by all out-of-field thoracic lymphoid organs are determined;• The target organ is left lung, radiation treatment planning (RTP) is such that in-field thoracic lymphoid organs comprise bone marrow, right lymph nodes (including LI, L2, L3 right axillary, sus-clavicular and interpectoral right lymph nodes), left lymph nodes (including LI, L2, L3 left axillary, sus-clavicular and interpectoral left lymph nodes), mammalian chain lymph nodes and mediastinal lymph nodes and out-of-field thoracic lymphoid organs comprise spleen and mesenteric lymph nodes and the irradiation doses received by all out-of-field thoracic lymphoid organs are determined; or• The target organ is esophagus, and radiation treatment planning (RTP) is such that in-field thoracic lymphoid organs comprise bone marrow, spleen, right axillary lymph nodes LI, L2, L3, left axillary lymph nodes LI, L2, L3, mammalian chain lymph nodes, mediastinal lymph nodes and out-of-field thoracic lymphoid organs comprise spleen, right sus-clavicular lymph nodes, right interpectoral lymph nodes, left sus-clavicular lymph nodes, left interpectoral lymph nodes and mesenteric lymph nodes and the irradiation doses received by all out-of-field thoracic lymphoid organs are determined.

11. The method of any one of claims 1-10, wherein said radiotherapy treatment is external radiotherapy, preferably 3D conformal radiation therapy (3DCRT), intensity-modulated radiation therapy (IMRT), volumetric modulated arc therapy (VMAT), stereotactic body radiation therapy (SBRT), image-guided radiation therapy (IGRT), including MRI-associated radiotherapy, proton therapy, high dose rate radiation therapy (HDRT) or ultra-high dose rate (UHDR) ; or internal radiotherapy or brachytherapy.

12. The method of any one of claims 1-11, for optimizing and / or personalizing a radiotherapy treatment plan to prevent radiation-induced lymphopenia.

13. The method of any one of claims 1-11, for selecting a radiotherapy treatment and / or optimizing prescription in order to minimize the detrimental effects of radiotherapy onto the immune system of said patient.

14. Use of the irradiation dose received by at least one out-of-field lymphoid organ during a radiotherapy treatment, in order to optimize and / or personalize said radiotherapy treatment to prevent radiation-induced lymphopenia, for implementing the method as defined in any one of claims 1-11.

15. A computer device configurated to implement the method as defined in any one of claims 1-11.

Images