Computer-implemented method for single time point dosimetry

Abstract

The present invention relates to a computer-implemented single time-point dosimetry method (200) comprising the following steps: a. Deriving for one or more organs at least one pre-therapy metrics from a pre-therapy scan of a patient (201); b. Deriving for the one or more organs the activity value at a specific time from a posttherapy scan of the patient (203); c. Computing the time-integrated activity value for the one or more organs (204); wherein the time-integrated value is computed based on the activity value from the posttherapy scan, the specific time of the posttherapy scan and a predicted effective half-life value, and wherein the predicted effective half-life value is determined by a machine learning-model trained according to the method of the present invention.

Description

[0001] Computer-implemented method for single time point dosimetry Technical Field The present invention relates to the field of dosimetry for radioligand therapy, in particular to the field of single time point dosimetry. In detail the present invention provides for a computer- implemented time-independent single time point dosimetry method using a trained machine learning-model and for a computer- implemented method for training the machine learning-model. Background of the invention Standard-of-care dosimetry is grounded on rigorous research and meticulous clinical trials, ensuring that patients undergo radiation treatments that are both effective and prioritize safety. Radiopharmaceutical therapy, and in particular Radioligand therapy (RLT), is still mostly limited to fixed injected activities with minimal treatment personalization. Furthermore, by actively practicing in only specialized centres, invaluable insights into the biodistribution of radioligands and subsequent tissue radiation dose post-administration are obtained. Such practice not only highlights the importance of dosimetry but also paves the way for enhanced patient care. By tailoring treatments to individual needs, it guides in a new era in personalized cancer management. In clinical settings, dosimetry often goes unapplied due to its inherent challenges. This valuable procedure can encounter challenges including limited clinical resources, patient comfortability, elevated imaging costs, scarce scanner availability and a time-consuming nature, thereby posing burdens for both patients and healthcare facilities, potentially escalating overall treatment costs. Nevertheless, substantial evidence underscores the profound impact of dosimetry and the tailoring of treatments in various therapeutic strategies. Estimating the radiation dose to organs or tissues upon RLT remains challenging. Factors like patient characteristics, tumour burden and the properties of the molecular compounds play a crucial role. However, the dosimetry is profoundly influenced by organ biokinetics and radioligand distribution. Information often sourced from multiple SPECT / CT scans over varied periods, termed as the multiple time point (MTP). This technique demands scans to ascertain dosimetry, based on organs and tumour biokinetics and radiopharmaceutical biodistribution, as well as the effective half-life (Teff) parameter. Researchers globally confront with the challenge of minimizing these scans to a singular point of measurement for cancer patients, i.e., single time point dosimetry (STP). From the state-of-art is known a single time point dosimetry method, the so-called Hänscheid method, to determine the time- integrated activity based on the activity measured from a posttherapy time scan. This method is however unpractical and does not allow to implement single time point dosimetry on a large scale due to the restriction of the measurement time point, particularly when patients are discharged on the same day treatment begins. Thus, there is a need for a method capable of providing accurate single time point dosimetry that address the limitations of conventional techniques and is early and flexible from the measurement time point. Summary of the invention Thus, the object of the present invention is to propose a novel single time point dosimetry method, with which the above-described drawbacks of the known methods are completely overcome or at least greatly diminished. According to the present invention, these objects are achieved in particular through the elements of the independent claims. Further advantageous embodiments follow moreover from the dependent claims and the description. In particular, the objects of the present invention are achieved by a computer-implemented method for training a machine learning- model to determine for one or more organs or tissues a patient specific predicted effective half-life (^^^^^^^^^^^^^^^^) value for radiopharmaceutical therapy, in particular radioligand therapy, the method comprising: a. Deriving at least one pre-therapy metrics from pre-therapy scans of patients for the one or more organs or tissues; b. Deriving for the one or more organs or tissues measured effective half-life values from a time series of posttherapy scans of the patients; c. Using the measured effective half-life values to train the machine learning-model to determine for the one or more organs or tissues a patient specific predicted effective half-life value from the at least one pre-therapy metrics. The objects of the present invention are also achieved by a computer-implemented single time point dosimetry method comprising the following steps: a. Deriving for one or more organs or tissues at least one pre- therapy metrics from a pre-therapy scan of a patient; b. Measuring at least one clinical value of the patient; c. Deriving for the one or more organs or tissues the activity value at a specific time from a posttherapy scan of the patient; d. Computing the time-integrated activity value for the one or more organs or tissues; characterized in that the time-integrated value is computed based on the activity value from the posttherapy scan, the specific time of the posttherapy scan and a predicted effective half-life value, wherein the predicted effective half-life value is determined by a machine learning-model trained according to the present invention. Thanks to the training method and the single time point dosimetry according to the present invention. It is possible to determine the time-integrated activity value for the one or more organs or tissues without having to apply the current gold standard method of multiple time point. With that it thus possible to reduce dramatically the number of scans that a patient must undergo during radioligand therapy. The present allows for reducing the costs of such therapies but also the time expenditure associated with such therapies. From the computed time- integrated activity value for the one or more organs or tissues, it is possible to compute the absorbed dose for said one or more organs or tissues. In a first preferred embodiment of the present invention, the machine learning-model is a random forest regressor or a convolutional neural network. A support vector regressor has the advantage that it provides a superior fit for the complex, high-dimensional data involved in training and using of the methods according to the present invention. Furthermore, support vector regressor delivers reliable and robust results, demonstrating strong performance even when applied to small datasets. However, other machine-learning and artificial intelligence models, especially for effective half-life or biokinetic prediction method, could be used. It can be also implemented in a voxel-wise way using a deep learning method. In a further preferred embodiment, the methods of the present invention comprise a step of measuring at least one pre-therapy blood value from pre-therapy blood samples of the patients, wherein the machine learning-model is trained to determine for the one or more organs or tissues a patient specific predicted effective half-life value from the at least one pre-therapy metrics and from the at least one pre- therapy blood value. Thanks to the at least blood value, the model reaches even better correspondence with the effective half-live value obtained by multiple time point approach. In a further preferred embodiment, the methods of the present invention comprise a step of measuring at least one clinical value of the patients, wherein the machine learning-model is trained to determine for the one or more organs or tissues a patient specific predicted effective half-life value from the at least one pre-therapy metrics and from the at least one clinical value of the patients. The pre-therapy scans and / or the posttherapy scans are advantageously PET / CT or SPECT / CT scans, wherein the theragnostic information of the imaging tracer is similar to the therapy compound. The at least one blood value is advantageously chosen from an organ- specific antigen value, for instance the prostate-specific antigen value, the lactate dehydrogenase value and the creatinine value. However, other blood sample values can add valuable significance to the model, such as haematocrit and haemoglobin. The at least one clinical value is advantageously chosen from the age, the weight and the sex of the patient(s) while the at least one pre-therapy metrics is advantageously chosen from the standard uptake volume mean of specific organs or tissues and the volume of specific organs or tissues. The posttherapy scans are advantageously acquired after injection of a radiopharmaceutical agent to the patient(s). In a further aspect, the present invention relates to a data processing device comprising means for carrying out the steps of the methods of the present invention. In another aspect, the present invention relates to a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the methods of the present invention. Brief description of the drawings Figure 1 shows a schematical workflow of a computer- implemented method for training a machine learning-model to determine a patient specific predicted effective half-life value for radiopharmaceutical therapy according to a preferred embodiment the present invention; Figure 2 shows a schematical workflow of a computer- implemented single time-point dosimetry method according to a preferred embodiment the present invention; Figure 3 shows an example of a simulated patient specific activity curves; Figure 4 shows relative absorbed dose differences (%) across all simulated data comparing both ML-based and Hänscheid single time point results against the multiple time point values; Figure 5 shows time scan dependency relative percentage differences of STP dosimetry (ML-based and Hänscheid) methods when comparing with MTP dosimetry for the simulated results; Figure 6 shows an empirical patient specific activity curve for the estimation of a patient specific measured effective half-life and time- integrated activity values; Figure 7 shows empirical relative absolute absorbed dose differences for the left kidney, right kidney, liver and spleen for the present method in comparison to the conventional Hänscheid single time point method; and Figure 8 illustrates empirical relative absolute differences grouped by time scan for the left kidney, right kidney, liver and spleen for the present method in comparison to the conventional Hänscheid single time point method. Detailed description of a preferred embodiment Figure 1 shows a schematical workflow of a preferred embodiment of a computer-implemented method 100 for training a machine learning-model to determine a patient specific predicted effective half-life value for radiopharmaceutical therapy according to the present invention. In a first step 101 at least one pre-therapy metrics is derived from pre-therapy scans of patients. The pre-therapy scans are advantageously PET scans and the at least one pre-therapy metrics is advantageously the mean standard uptake volume of specific organs or tissues or the volume of specific organs or tissues. In a second step 102 at least one clinical value of the patients is measured. Advantageously, here at least one pre-therapy blood value is also measured from pre-therapy blood samples of the patients. The blood value is advantageously an organ-specific antigen value, for instance the prostate-specific antigen value, or the lactate dehydrogenase value or the creatinine value. The clinical value is for instance the weight, age or sex of the patients. In a third step 103 measured effective half-life values are derived from a time series of posttherapy scans of the patients. Deriving effective half-life values from a series of posttherapy scans is a well- known procedure in the field of multiple time-points dosimetry. In a fourth step 104 the measured effective half-life values are used to train the machine learning-model to determine a patient specific predicted effective half-life value from the at least one pre- therapy metrics, the at least one pre-therapy blood value, if present, and the at least one clinical value. The machine learning-model is advantageously a random forest regressor or a support vector regressor. Figure 2 shows a schematical workflow of a preferred embodiment of a computer-implemented single time-point dosimetry method 200 according to the present invention. In a first step 201 at least one pre-therapy metrics is derived from a pre-therapy scan of a patient. The pre-therapy scans are advantageously PET scans and the at least one pre-therapy metrics is advantageously the mean standard uptake volume of specific organs or tissues or the volume of specific organs or tissues. In a second step 202 at least one clinical value of the patient, and advantageously at least one pre-therapy blood value is measured from a pre-therapy blood sample of the patient, is measured. The blood value is advantageously an organ-specific antigen value, for instance the prostate-specific antigen value, or the lactate dehydrogenase value or the creatinine value. The clinical value is for instance the weight, age or sex of the patients. In a third step 203 the activity value for one or more organs or tissues is derived at a specific time from a posttherapy scan of the patient. In a fourth step 204 the time-integrated activity value for the one or more organs or tissues is computed based on the activity value from the posttherapy scan, the specific time of the posttherapy scan and the predicted effective half-life value obtained for the patient by means of the machine learning-model trained by means of the method according to the present invention. In step 205, the absorbed dose for said one or more organs or tissues is calculated from the computed time-integrated activity value for the one or more organs or tissues by means of equation (4). Concrete details of implementation of the training of the present machine learning model and of the implement used of this trained model for single time-point dosimetry are disclosed below in the Results section. Results I. Simulations The methods of the present invention were first validated by using fifty simulated patients in the proof-of-concept study below. Data of forty-five patients were used for the training of the machine learning- model while the data of five patients were used for the validation. The mean weight of the patients was chosen to be 71 kg with a standard deviation of ±16 kg. Pre-therapy scans were simulated using an injected activity dose of 112.1 ± 16.0 MBq of [177Lu]Lu-PSMA I&T for the following organs: heart, kidneys, liver, lungs, muscle, pancreas, salivary glands, spleen and thyroid. From these simulated pre-therapy scans, the standard uptake value for each of the organs were derived by collecting the volume of the organs. Posttherapy scans were simulated using an injected activity dose of 7.4 ± 0.3 GBq of [177Lu]Lu-PSMA I&T for the following organs: kidney, liver, salivary glands and spleen. Patient specific measured effective half-life values were estimated from patient specific activity curves (see Figure 3 for an example) considering simulated time scans at 2h, 20h, 43h, 69h, 144h, 165h and 192h. To model the pharmacokinetic of [177Lu]Lu-PSMA I&T separately in kidneys, liver, salivary glands and spleen, the following exponential equation was used to account for an initial uptake and rapid clearance phases: where ^^^^1and ^^^^2are the measured organ activities (MBq), and ^^^^2are the biological decay constants, and ^^^^^^^^ℎ^^^^^^^^is the physical decay constant of177Lu (6.65 days). The measured effective half-life (^^^^^^^^^^^^^^^^) and the time- integrated activity (TIA), further represented by Ã, were calculated. Two established approaches have been selected to estimate the time-integrated activity (Ã) from STP (single time-point) data: i) corresponding to activity uptake after a p.i. ^^^^^^^^^^^^, characterized by the predicted ^^^^^^^^^^^^^^^^(Equation 2); and ii) based on the conventional method developed by Hänscheid (Equation 3) [Hänscheid et al. Dose mapping after endoradiotherapy with 177Lu-DOTATATE / DOTATOC by a single measurement after 4 days. J Nucl Med.2018;59(1):75-81; doi: 10.2967 / jnumed.117.193706.]. Note that the time scan ^^^^^^^^^^^^references the specific time when the scan for the single time-point dosimetry was conducted. The STP methods are henceforth referred to as i) machine learning based (ML-based), as the ^^^^^^^^^^^^^^^^predicted by the ML model was utilized, and ii) Hänscheid. Dosimetry was performed for the kidneys, liver, salivary glands and spleen using the ML-based and Hänscheid methods. The absorbed dose (^^^^(^^^^^^^^,^^^^^^^^)) is given by the product between the TIA (Ã(^^^^^^^^,^^^^^^^^)) andthe absorbed dose per decay (^^^^(^^^^^^^^ ← ^^^^^^^^)), in ^^^^^^^^^^^^ / ^^^^^^^^^^^^ ∙ ^^^^

[0017] : The S-values (^^^^) were obtained from the webpage OpenDose (www.opendose.org) and the units were converted resulting the absorbed dose in Gy. For each specific organ (^^^^) (kidneys, liver, salivary glands or spleen), a data matrix was defined. In this matrix, individual patients are represented by rows, while their corresponding features form the columns. The feature selection (^^^^) encompasses not only basic patient characteristics like weight but also comprise pre-therapy PET / CT metrics such as the specific organ per SUV ratio and volume. Additionally, the matrix incorporates other pertinent clinical data, all of which is curated and structured according to the ^^^^^^^^^^^^^^^^function: The machine learning model employed is a RandomForestRegressor (RFR), as ensemble learning method for regression for constructing a multitude of decision trees during training and generating the mean probability of the individual trees. The RFR was chosen here because it provides a superior fit for the complex, high- dimensional data involved in this study. The model validation was conducted using k-fold cross-validation. The data set was partitioned into ^^^^ = 10 folds. The error function for performance evaluation wascomputed as the mean percentage error between the predicted (simulated) and measured effective half-life, (a lower mean error indicated a superior predictive organ model). Ultimately, the performance was summarized by calculating the mean and standard deviation of the mean error across the folds, providing an aggregate view of the model's performance. The relative absorbed dose differences (%) were calculated using the following equation:100 (6) where, ^^^^^^^^^^^^^^^^is the absorbed dose from the MTP dosimetry and ^^^^^^^^^^^^^^^^calculated from the STP dosimetry. The STP dosimetry was conducted utilizing Eq.4, wherein the respective à was computed employing two distinct methods: by applying the ^^^^^^^^^^^^^^^^predicted (simulated) from the ML-model (Eq.2) the Hänscheid method (Eq.3). The ^^^^^^^^^^^^^^^^results are shown in Table 1 for kidney, liver, salivary glands and spleen. Organ Time ML Based Model Hänscheid Method Scan RDD (%) RDD (%) (h) 2 9.6 ± 0.6 -86.0 ± 0.1 20 8.7 ± 0.6 -8.7 ± 0.5 43 10.4 ± 0.2 17.2 ± 0.2 Kidneys 69 25.1 ± 0.2 16.9 ± 0.2 144 83.2 ± 0.2 -37.0 ± 0.1 165 103.6 ± 0.2 -50.6 ± 0.1 192 133.1 ± 0.2 -64.8 ± 0.0 2 106.6 ± 0.7 -70.2 ± 0.1 20 -14.0 ± 0.5 -22.9 ± 0.4 43 -18.6 ± 0.3 -14.3 ± 0.3 Liver 69 -1.4 ± 0.3 -15.9 ± 0.3 144 82.8 ± 0.5 -54.8 ± 0.1 165 117.2 ± 0.6 -64.6 ± 0.1 192 170.9 ± 0.7 -74.7 ± 0.1 2 -4 ± 0.6 -86.2 ± 0.1 20 16.2 ± 0.6 4.2 ± 0.6 Salivary glands 43 19.9 ± 0.2 26.3 ± 0.2 69 33.3 ± 0.2 13.7 ± 0.2 144 85.4 ± 0.2 -54.0 ± 0.1 165 103.2 ± 0.2 -66.8 ± 0.0 192 128.3 ± 0.2 -78.6 ± 0.0 2 60.8 ± 0.6 -78.0 ± 0.1 20 -6.2 ± 0.5 -18.0 ± 0.5 43 -7.1 ± 0.2 -1.6 ± 0.3 Spleen 69 9.7 ± 0.3 -2.6 ± 0.3 144 83.9 ± 0.4 -47.5 ± 0.1 165 112.4 ± 0.5 -58.9 ± 0.1 192 155.4 ± 0.5 -70.7 ± 0.1 Table 1: Teffprediction for the kidneys, liver, salivary glands and spleen. Values presented include mean ± standard deviation values. The discrepancies in the absorbed doses, presented as relative absorbed dose differences (RDD) (%) per organ, were benchmarked against the gold standard MTP dosimetric values using Eq.6. A graphical representation of these differences across the VOIs are given in Figure 4. In the present study, Hänscheid method had a small variance in organ differences when contrasted to the results from the ML-based method. Specifically, for the kidneys, we observed a mean relative difference of -30.4 ± 37.3%. This was closely followed by differences of - 45.3 ± 24.8%, -34.5 ± 44.1%, and -39.6 ± 29.7% for the liver, salivary glands, and spleen, respectively. On the other hand, when employing ML-based model to calculate the absorbed dose, the RDD for these organs were: 53.4 ± 48.4% for the kidneys, 63.4 ± 69.4% for the liver, 54.6 ± 46.8% for the salivary glands, and 58.4 ± 58.4% for the spleen. The disparities between the two methods, Hänscheid and ML-based, were not merely observational. Furthermore, to determine the statistical significance of these differences, a Wilcoxon test was performed, indicating a highly statistically significant difference (p<0.001). The RDD (%) calculated from the ML-based and Hänscheid methods were treated against the MTP organized per organ. Moreover, to assess the robustness of the present ML-model, the RDD were separated by time ^^^^^^^^^^^^. The results of this analysis focus on the statistical relationship between the methods over multiple ^^^^^^^^^^^^for the kidneys, liver, and spleen. Figure 5 presents the mean and standard deviation of the RDD values, and statistical significances comparing them from the STP methods against MTP for the kidneys, liver, salivary glands and spleen. We evaluated all ^^^^^^^^^^^^to verify the variation in RDD values across different scanning intervals between the ML based model and Hänscheid method, as shown in Figure 5. Analysis revealed significant disparities in RDD across all ^^^^^^^^^^^^, with marked percentage variations for the kidneys, liver, salivary glands and spleen over 80% at 144 h ^^^^^^^^^^^^and over 100% at 165 h and 192 h ^^^^^^^^^^^^, when using ML-based model, including 2 h ^^^^^^^^^^^^for the liver. This trend was contrasted with the results derived from the Hänscheid method at same time scans, showing the highest mean RDD value of -86.2 ± 0.1% for the salivary glands at 2 h ^^^^^^^^^^^^. Additionally, at other scan times, it is evident that either the machine learning-based model effectively reduces the RDD values, or the Hänscheid model does.At the described ^^^^^^^^^^^^, statistical significances were observed (^^^^ < 0.0001).Contrasting with the Hänscheid method, our ML-based approach exhibited the minimal mean RDD for the kidneys at 2 h and 43 h ^^^^^^^^^^^^, revealing -9.6 ± 0.6% and 10.4 ± 0.2%, versus -86.0 ± 0.1% and 17.2 ± 0.2%, respectively. Notably, at 20 h, both methods exhibited nearly identical RDD values. Statistical testing conducted for these ^^^^^^^^^^^^showedsignificance, with ^^^^ < 0.001. Conversely, at 69 h and 144 h, theHänscheid method (16.9 ± 0.2% and -37.0 ± 0.1%) proved more effective, yielding lower RDD values compared to the ML-based model (25.1 ± 0.2% and 83.2 ± 0.2%). For the liver model, our method revealed a minimum mean RDD of -14.0 ± 0.5% at 20 h and -1.4 ± 0.3% at 69 h. In comparison, at the same time scan, the Hänscheid method presented -22.9 ± 0.4% and -15.9 ± 0.3%, respectively. At 43 h, a slightly difference (4.3%) was observed favorable to Hänscheid method. Across the mentioned ^^^^^^^^^^^^,statistical significances were found (^^^^ < 0.0001). When assessing the RDDvalues for the salivary glands between the ML-based model and the Hänscheid method, minimal differences were observed at 2 h, 43 h, and 69 h time scans. However, at the 20 h ^^^^^^^^^^^^, the RDD was found to be 4.2 ± 0.6% for the Hänscheid method as opposed to 16.2 ± 0.6% for the ML- based model. The RDD calculated for the spleen from the ML-based model presented a notable minimum mean variance at 2 h and 20 h ^^^^^^^^^^^^, being 60.8 ± 0.6% and -6.2 ± 0.5%, respectively. In contrast, Hänscheid method, -78.0 ± 0.1% and -18.0 ± 0.5%, for the same time scans. At 43 h ^^^^^^^^^^^^, RDD results from both methods exhibited differences less than 10%.Statistic significances with ^^^^ < 0.0001 were found across all ^^^^^^^^^^^^.II. Empirical results The methods of the present invention were further validated by using empirical results. Fifteen patients with metastatic castration- resistant prostate cancer (mCRPC) underwent [177Lu]Lu-PSMA 617 radioligand therapy at Department of Nuclear Medicine, Inselspital, Bern, Switzerland. A total of 55 therapy cycles were included, comprising pre-therapy [18F]F-PSMA-1007 PET / CT scans before first therapy cycle. To model the pharmacokinetic of [177Lu]Lu-PSMA 617 separately in left kidney, right kidney, liver, and spleen, the following exponential equation was used to account for an initial uptake and rapid clearance phases: where A is the measured organ activity and ^^^^2are the biological decay constants, and ^^^^^^^^ℎ^^^^^^^^is the physical decay constant of 177Lu (6.65 days). The measured effective half-life (^^^^^^^^^^^^^^^^) and the time- integrated activity (TIA), further represented by Ã, were calculated. The parameters ^^^^ and ^^^^2were considered as population data to fit the following function: where ^^^^1is the individual activity measured at a time scan. Furthermore, assuming ^^^^′(^^^^) as mono-exponential function, Equation 8 was fitted to cycles with at least three post-therapeutic SPECTtime points to estimate the effective half-life (^^^^^^^^^^^^^^^^ = ln(2 )and time-integrated activity (TIA). Two established approaches have been selected to estimate the time-integrated activity (Ã) from STP (single time-point) data: i) corresponding to activity uptake after a p.i. ^^^^^^^^^^^^, characterized by the predicted ^^^^^^^^^^^^^^^^(Equation 8); and ii) based on the conventional method developed by Hänscheid (Equation 9) [Hänscheid et al. Dose mapping after endoradiotherapy with 177Lu-DOTATATE / DOTATOC by a single measurement after 4 days. J Nucl Med.2018;59(1):75-81; doi: 10.2967 / jnumed.117.193706.]. Note that the time scan ^^^^^^^^^^^^references the specific time when the scan for the single time-point dosimetry was conducted. The STP methods are henceforth referred to as i) instant single time point (iSTP), as the ^^^^^^^^^^^^^^^^predicted by the ML model was utilized, and ii) Hänscheid. Dosimetry was performed for the left kidney, right kidney, liver, and spleen using the ML-based and Hänscheid methods. The absorbed dose (^^^^(^^^^^^^^,^^^^^^^^)) is given by the product between the TIA (Ã(^^^^^^^^,^^^^^^^^)) andthe absorbed dose per decay (^^^^(^^^^^^^^ ← ^^^^^^^^)), in ^^^^^^^^^^^^ / ^^^^^^^^^^^^ ∙ ^^^^

[0017] : The S-values (^^^^) were obtained from the webpage OpenDose and the units were converted resulting the absorbed dose in Gy. For each specific organ (^^^^) (left kidney, right kidney, liver, and spleen), a data matrix was defined. In this matrix, individual patients are represented by rows, while their corresponding features form the columns. The feature selection (^^^^^^^^) encompasses not only basic patient characteristics like weight but also comprise pre-therapy PET / CT metrics such as the specific organ SUV mean. Additionally, the matrix incorporates other pertinent clinical data, all of which is curated and structured according to the ^^^^^^^^^^^^^^^^function: The machine learning model employed here is a SupportVectorRegressor (SVR), as employing a radial basis function kernel to construct a hyperplane in high-dimensional space while minimizing error and controlling model complexity. The SVR was chosen because it provided a superior fit for the small datasets involved in this empirical study. The model validation was conducted using k-fold cross-validation. The data set was partitioned into ^^^^ = 10 folds. The errorfunction for performance evaluation was computed as the mean percentage error between the predicted (simulated) and measured effective half-life, (a lower mean error indicated a superior predictive organ model). Ultimately, the performance was summarized by calculating the mean and standard deviation of the mean error across the folds, providing an aggregate view of the model's performance. The relative absolute differences (%), in time-integrated activity, were calculated using the following equation: 100 (11) where, Τ^^^^^^^^^^^^is the time-integrated activity from the MTP dosimetry and Τ^^^^^^^^^^^^calculated from the STP methods. As mentioned above, a machine learning (ML)-model was used for ^^^^^^^^^^^^^^^^predictions for the left and right kidney, liver and spleen, subsequently used to estimate time-integrated activity (TIA) and absorbed dose. A total of 55 patient-cycle-wise were used to train and test the machine learning-model, here a Support Vector Regressor (SVR), applying 10-fold cross-validation. The results of the present methods were compared against the gold standard for dosimetry MTP and the previously proposed STP Hänscheid methods. In the here presented implementation, the methods comprised patient pre-therapy data before the first cycle as features to predict ^^^^^^^^^^^^^^^^. In other words, the pre- therapy data obtained before the first cycle were used to predict ^^^^^^^^^^^^^^^^for the first cycle but also for the subsequent cycles. The present method based on SVR model yields to predicted organ’s ^^^^^^^^^^^^^^^^with a mean error (ME) of 23.6 ± 5.9% for the left kidney, 21.3 ± 6.4% for the right kidney, 20.3 ± 10.4% for the liver and 28.4 ± 10.2% for the spleen compared to MTP, as summarized in Table 2. Support Vector Regressor Organ^^^^^^^^^^^^^^^^ Mean Error (%)Left kidney 23.6 ± 5.9 Right kidney 21.3 ± 6.4 Liver 20.3 ± 10.4 Spleen 28.4 ± 10.2 Table 2: ^^^^^^^^^^^^^^^^mean error resulting from the trained Support Vector Regressor model MTP was used to estimate the absorbed doses and ^^^^^^^^^^^^^^^^results that served as reference gold standard (see Table 3). The values of Table 3 were used for the evaluation of different STP methods, as shown in Table 4. Posttherapy scans were acquired using an injected activity dose of 7.4 ± 0.3 GBq of [177Lu]Lu-PSMA 617 for the following organs: left kidney, right kidney, liver and spleen. Patient specific measured effective half-life values were estimated from patient specific activity curves (see Figure 6 for an example for the left kidney) considering simulated time scans at 2 h, 24 h, 48 h, 72 h, 192 h and 216h. Organ Absorbed Dose Mean ± SD (Gy) Effective Half-Life Mean ± SD (h) Left kidney 2.34 ± 2.34 38 ± 37 Right kidney 2.52 ± 2.58 41 ± 44 Liver 0.62 ± 0.53 26 ± 14 Spleen 1.36 ± 5.21 21 ± 20 Table 3 Reference values from the multiple time point method of the absorbed dose and effective half-life for the left kidney, right kidney, liver and spleen. The instant single time point dosimetry was derived from the predicted effective half-life values. The present method results were compared to the standard multiple time point (MTP) method, which is based on the relative absolute time-integrated activity differences (%). Additionally, the empirical results were compared to the conventional STP method proposed by Hänscheid et al. (Hänscheid H, Lapa C, Buck AK, Lassmann M, Werner RA. Dose mapping after endoradiotherapy with 177Lu-DOTATATE / DOTATOC by a single measurement after 4 days. J Nucl Med. 2018;59:75-81). The mean relative absolute differences in absorbed dose estimated from each method (where iSTP stands for the single time point dosimetry method of the present invention) are plotted in Figure 7 for each organ. The iSTP method yielded mean differences of 15.94 ± 13.31% for the left kidney, 14.76 ± 14.05% for the right kidney, 17.82 ± 22.05 % for the liver and 39.06 ± 50.68% for the spleen. The p-values for the comparisons with the Hänscheid method indicated statistical significance for the left kidney (p < 0.001) and the liver (p < 0.05), while the comparisons for the right kidney (p = 0.074) and the spleen (p = 0.210) were not statistically significant (See Figure 7). Table 4 summarizes the achievements by time scan (2 h, 24 h, 48 h and 72 h), while Figure 8 illustrates the results for the left kidney, right kidney, liver and spleen. Organ Time scan (h) iSTP Hänscheid Wilcoxon P- value Left kidney 2 23.2 ± 15.9 87.2 ± 3.9 0.00 (***) 24 11.0 ± 7.9 12.0 ± 8.8 0.30 (ns) 48 11.0 ± 8.1 9.3 ± 7.4 0.00 (**) 72 21.5 ± 14.3 13.6 ± 9.2 0.00 (***) Right kidney 2 20.7 ± 15.6 88.1 ± 3.1 0.00 (***) 24 12.4 ± 10.9 12.9 ± 10.5 0.45 (ns) 48 10.1 ± 11.9 8.1 ± 11.3 0.03 (*) 72 16.7 ± 13.2 17.0 ± 13.7 0.96 (ns) Liver 2 18.9 ± 23.0 83.4 ± 5.8 0.00 (***) 24 12.5 ± 15.8 10.9 ± 12.8 0.63 (ns) 48 12.3 ± 11.6 11.0 ± 9.3 0.99 (ns) 72 38.2 ± 29.8 20.6 ± 16.5 0.01 (**) Spleen 2 25.5 ± 23.1 77.2 ± 8.3 0.00 (***) 24 18.2 ± 20.7 16.0 ± 19.2 0.00 (***) 48 40.0 ± 41.6 26.0 ± 21.6 0.01(*) 72 133.5 ± 125.8 32.3 ± 18.5 0.00 (***) Table 4: Relative absolute differences (RAD) (%) in time-integrated activity using instant single time point and Hänscheid method for left and right kidneys, liver and spleen by time scan (h) (where *, ** and *** refers to significant differences comparing the methods, and ’ns’ means no significant differences). Data from left kidney, right kidney, liver and spleen were analyzed using the iSTP method applied during the early time after injection points: 2h, 24 h, 48 h and 72 h. Hänscheid method performs better when the time point scan fells in the interval 0.75 and 2.5 times ^^^^^^^^^^^^^^^^. For this specific method, we selected 24 h, 48 h and 72 h ^^^^^^^^^^^^, based on the obtained mean value of ^^^^^^^^^^^^^^^^from the MTP, which achieved best optimal results for the kidney, salivary glands, liver and spleen. Specifically, RDD below 17% for both kidneys and 20% for the liver. For the left kidney, particularly at 2 h ^^^^^^^^^^^^, the iSTP method showed a mean RAD of 23.18 ± 15.94%, with 11.02 ± 8.05% at 24 h. Furthermore, lower errors were recorded at 48 h (11.02 ± 8.05%) and 72 h (21.45 ± 14.29%) ^^^^^^^^^^^^. For the right kidney, at 2 h ^^^^^^^^^^^^, a mean difference of 20.66 ± 15.57% was noted. The RAD errors at 24 h, 48 h and 72 h were 12.35 ± 10.92%, 10.05 ± 11.85% and 16.73 ± 13.17%, respectively. Regarding the liver, this method exhibited a RAD of 18.87 ± 22.97% at 2 h. At 24 h ^^^^^^^^^^^^, a lower RAD of 12.45 ± 15.76% was observed, followed by RAD values of 12.27 ± 11.63% at 48 h and 38.22 ± 29.83% at 72 h ^^^^^^^^^^^^. In the analysis comparing the iSTP and Hänscheid methods across 24 h, 48 h and 72 h time points, no statistically significant differences (>0.05) were found at 24 h for both kidneys and liver, 48 h for the liver and 72 h for the right kidney, except for the spleen. Slight significances were observed at 48 h for the left and right kidney (p<0.05) between both methods. The empirical results above show that the proposed instant iSTP method streamlines dosimetry and offers flexibility in selecting early post- treatment imaging time points, specifically at 2 or 4 h and 24 h. The findings above show furthermore that, despite variations in relative absolute differences, the present iSTP method demonstrated the ability to predict the effective half-life using pre-therapy data, enabling STP to be conducted shortly after radioligand therapy (RLT).

Claims

Claims 1. A computer-implemented method for training a machine learning-model to determine for one or more organs or tissues a patient specific predicted effective half-life value for radiopharmaceutical therapy (100), in particular radioligand therapy, the method comprising: a. Deriving at least one pre-therapy metrics from pre-therapy scans of patients for the one or more organs or tissues (101); b. Deriving for the one or more organs or tissues measured effective half-life values from a time series of posttherapy scans of the patients (103); c. Using the measured effective half-life values to train the machine learning-model to determine for the one or more organs or tissues a patient specific predicted effective half-life value from the at least one pre-therapy metrics.

2. Computer-implemented single time-point dosimetry method (200) comprising the following steps: a. Deriving for one or more organs at least one pre-therapy metrics from a pre-therapy scan of a patient (201); b. Measuring at least one clinical value of the patient (202); c. Deriving for the one or more organs or tissues the activity value at a specific time from a posttherapy scan of the patient (203); d. Computing the time-integrated activity value for the one or more organs or tissues (204); characterized in that the time-integrated value is computed based on the activity value from the posttherapy scan, the specific timeof the posttherapy scan and a predicted effective half-life value, wherein the predicted effective half-life value is determined by a machine learning-model trained according to claim 1.

3. The methods according to claim 1 or 2 wherein the machine learning-model is a random forest regressor, a support vector regressor or a convolutional neural network.

4. The methods according to any of the preceding claims comprising a step of measuring at least one pre-therapy blood value from pre-therapy blood samples of the patients, wherein the machine learning-model is trained to determine for the one or more organs or tissues a patient specific predicted effective half-life value from the at least one pre-therapy metrics and from the at least one pre-therapy blood value.

5. The methods according to any of the preceding claims comprising a step of measuring at least one clinical value of the patients (102), wherein the machine learning-model is trained to determine for the one or more organs or tissues a patient specific predicted effective half-life value from the at least one pre-therapy metrics and from the at least one clinical value of the patients (102).

6. The methods according to claim any of the preceding claims, wherein the pre-therapy scans and / or the posttherapy scans are PET / CT or SPECT- / CT scans, wherein the theragnostic information of the imaging tracer is advantageously similar to the therapy compound.

7. The methods according to any of the preceding claims, wherein the at least one blood value is chosen from an organ-specific antigen value, for instance the prostate-specific antigen value, the lactate dehydrogenase value and the creatinine value.

8. The methods according to any of the preceding claims wherein, the at least one clinical value is chosen from the age, the weight and the sex of the patient(s).

9. The methods according to any of the preceding claims wherein, the at least one pre-therapy metrics is chosen from the standard uptake volume mean of specific organs and the volume of specific organs.

10. The methods according to any of the preceding claims wherein the posttherapy scans are acquired after injection of a radiopharmaceutical agent to the patient(s).

11. The method according to claim 2, wherein the activity value is derived from a posttherapy scan of the patient acquired at a time between 2-4h or between 20-24h posttherapy.

12. A data processing device comprising means for carrying out the steps of the method of claim 1 and / or the steps of the method of claim 2.

13. A computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method of claim 1 and / or the steps of the method of claim 2.

14. A computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method of claim 1 and / or the steps of the method of claim 2.

Images