Method for predicting the tumor stage of a subject in a treatment phase and associated devices

Abstract

The prediction of the tumor stage is very difficult and is notably problematic to carry out a proper follow-up of the treatment of the tumor. The Applicant has found a way to circumvent such problem by proposing a computer tool which is adapted to predict the tumor stage of a subject, based on original parameters; a volume of the tumor calculated based on images obtained by one modality, at least one parameter relative to the interaction between a chemical agent and the tumor, and t least one parameter relative to the presence of tumor vessels in a volume containing the tumor.

Description

[0001] METHOD FOR PREDICTING THE TUMOR STAGE OF A SUBJECT IN A TREATMENT PHASE AND ASSOCIATED DEVICES

[0002] TECHNICAL FIELD OF THE INVENTION

[0003] The present invention concerns a method for predicting the tumor stage of a subject in a treatment phase. The present invention also concerns associated devices, namely a computer program product and a support.

[0004] BACKGROUND OF THE INVENTION

[0005] Monitoring the response to therapy is important for precision oncology.

[0006] In 1979, the World Health Organization (WHO) determined four categories of response or non-response to treatment based on the tumor volume: Complete, Partial, Stable and Progressive Disease. In 2000, the Response Evaluation Criteria in Solid Tumors (RECIST) proposed to sum one-dimensional measurements of the longest length of all lesions. In 2009, Positron Emission Tomography (PET) Response Criteria in Solid Tumors (PERCIST) introduced a continuous variable, the percentage change between pre- and post-treatment PET scans of the lean-body-mass-corrected peak Standard Uptake Value (SUL), or of the sum of all SULs of all lesions. The RECIST and PERCIST criteria inform on gross characteristics of tumors, but provide little if any information about the action of a personalized treatment on its pharmacological target, e.g., immune checkpoint inhibition, anti-angiogenesis, targeted immunotherapy, etc... The limited capacity of RECIST and PERCIST to report on the action of a new treatment on its specific target stands in contrast with the increasing availability of in vivo molecular and functional imaging approaches targeting tumor hallmarks.

[0007] Paragangliomas (PGLs) are neuroendocrine tumors arising from extra-adrenal chromaffin cells that originate from the neural crest cells and are characterized by high metabolism and high vascularization. Genetic studies have reported that a large number of PGL carry germline mutations of several genes. PGL are often benign but mutations in the SDHB gene coding for the B-subunit of succinate dehydrogenase (SDHB) lead to a malignant form of the disease with multiple metastasis and a low survival rate. Sunitinib is an anti-angiogenic drug under trial for malignant PGL patients, however at least one third of SDHB mutation carriers escape sunitinib treatment. SUMMARY OF THE INVENTION

[0008] There is therefore a need for a method for predicting the tumor stage of a subject with a good accuracy.

[0009] To this end, the specification describes a method for predicting the tumor stage of a tumor of a subject, the method being computer-implemented, the method comprising the steps of :

[0010] - providing images of the tumor of the subject obtained by several modalities,

[0011] - extracting a set of parameters from the provided images, the set of parameters comprising:

[0012] - a volume of the tumor calculated based on images obtained by one modality,

[0013] - at least one parameter relative to the interaction between a chemical agent and the tumor, and

[0014] - at least one parameter relative to the presence of tumor vessels in a volume containing the tumor, and

[0015] - determining the tumor stage based on the extracted set of parameters.

[0016] According to further aspects of the method for predicting, which are advantageous but not compulsory, the method for predicting might incorporate one or several of the following features, taken in any technically admissible combination:

[0017] - the extracted set of parameters comprises several values for volume of the tumor, each value being obtained by calculation on images obtained by a respective modality.

[0018] - the several modalities at the step of providing comprises ultrasound imaging and positron emission tomography.

[0019] - the several modalities at the step of providing consists of positron emission tomography, ultrasound imaging and computed tomography imaging.

[0020] - the at least one parameter relative to the interaction between a chemical agent and the tumor is the concentration of the chemical agent in a volume containing the tumor.

[0021] - the at least one parameter relative to the presence of tumor vessels is the number of nodes, the density of nodes or a length of at least one vessel.

[0022] - the at least one parameter relative to the length is chosen in the list consisting of :

[0023] - the average of the minimum length of each vessel,

[0024] - the average of the minimum length of each vessel, and

[0025] - the variance of the mean vessel lengths.

[0026] - the set of parameters consists in:

[0027] - a volume of the tumor calculated based on images obtained by several modalities, - one parameter relative to the concentration of a chemical agent in a volume containing the tumor, and

[0028] - one parameter relative to the presence of tumor vessels in a volume containing the tumor.

[0029] - each stage is associated to a domain of values for each parameter and, during the step of determining, the stage is determined based on the domain to which the extracted set of parameters belongs.

[0030] - the number of stages is equal to 4 and wherein the first stage corresponds to an early stage of the tumor, - the second stage corresponds to a first intermediate reversible stage of the tumor, - the third stage corresponds to a second intermediate reversible stage of the tumor, and - the fourth stage corresponds to an irreversible end stage of the tumor.

[0031] - the tumor is a paraganglioma.

[0032] - the subject is a mammal.

[0033] The specification also concerns a computer program product comprising instructions for carrying out the steps of a method as previously described when said computer program product is executed on a suitable computer device.

[0034] The specification also deals with a computer readable medium having encoded thereon a computer program product.

[0035] BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The invention will be better understood on the basis of the following description which is given in correspondence with the annexed figures and as an illustrative example, without restricting the object of the invention. In the annexed figures:

[0037] - figure 1 shows schematically a system and a computer program product which interaction enables to carry out a method for predicting the tumor stage of a subject in a treatment phase,

[0038] - figure 2 is a flowchart illustrating a specific example of carrying out a method for predicting the tumor stage of a subject in a treatment phase,

[0039] - figure 3 is a flowchart illustrating a pipeline enabling to obtain the parameters involved in the method for predicting,

[0040] - figure 4 illustrates schematically the database generation process, and

[0041] - figure 5 illustrates the possible evolution between the tumor stages. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0042] A system 10 and a computer program product 12 are represented in figure 1. The interaction between the computer program product 12 and the system 10 enables to carry out a method for predicting the tumor stage of a subject in a treatment phase.

[0043] The method for predicting is thus a computer-implemented method.

[0044] System 10 is a computer. In the present case, system 10 is a laptop.

[0045] More generally, system 10 is a computer or computing system, or similar electronic computing device adapted to manipulate and / or transform data represented as electronic quantities within the computing system's registers and / or memories into other data similarly represented as physical quantities within the computing system's memories, registers or other such information storage, transmission or display devices.

[0046] System 10 comprises a processor 14, a keyboard 22 and a display unit 24.

[0047] The processor 14 comprises a data-processing unit 16, memories 18 and a reader 20. The reader 20 is adapted to read a computer readable medium.

[0048] The computer program product 12 comprises a computer readable medium.

[0049] The computer readable medium is a medium that can be read by the reader of the processor. The computer readable medium is a medium suitable for storing electronic instructions, and capable of being coupled to a computer system bus.

[0050] Such computer readable storage media are, for instance, disks, floppy disks, optical disks, CD-ROMs, magnetic-optical disks, read-only memories (ROMs), random access memories (RAMs), electrically programmable read-only memories (EPROMs), electrically erasable and programmable read only memories (EEPROMs), magnetic or optical cards, or any other type of medium for storing electronic instructions, and able to be coupled to a computer system bus.

[0051] A computer program is stored in the computer readable storage medium. The computer program comprises one or more stored sequences of program instructions.

[0052] The computer program is loadable into the data-processing unit and adapted to cause execution of the method for predicting when the computer program is run by the data- processing unit.

[0053] Operation of the system 10 is now described in reference to the flowchart of figure 2, which illustrates an example of carrying out the method for predicting.

[0054] The method for predicting aims at predicting the tumor stage of a subject.

[0055] A tumor is a neoformation of body tissue (neoplasia) that occurs as a result of a disturbance in cell growth. A tumor is either benign or malignant, a malignant tumor is called cancer. The method for predicting applies to malignant tumors located in sites that can be detected using medical imaging, including, but not limited to, neuroendocrine tumors, renal cell carcinoma, hepatocellular carcinoma, and other tumors in solid organs.

[0056] The method for predicting is particularly useful in tumors that are treated by drugs or external physical treatments such as radiotherapy.

[0057] According a specific example, the tumor is a paraganglioma.

[0058] Paragangliomas are rare neuroendocrine tumors arising from chromaffin cells that originate from the neural crest cells and are located in one or several ganglia of the sympathetic and parasympathetic chains extending from the base of the skull to the pelvis.

[0059] The subject is a mammal, preferably a human or a rodent.

[0060] The method is preferably carried during the treatment phase, meaning that the images are taken at different moment of the treatment phase.

[0061] The method can also be applied during or in between treatment cycles. It is also useful for follow-up after treatment periods or during remission.

[0062] The method for predicting comprises a step of providing, a step of extracting and a step of determining.

[0063] During the step of providing, images of the tumor of the subject obtained by several modalities are provided to the system 10.

[0064] The images are either taken by one or several imaging systems which is part of the system 10 or are received from such a system 10.

[0065] In the present case, the images of a tumor are obtained by a multi-modal Imaging system and this multi-modal Imaging system is not included in the system 10.

[0066] As a specific example, the multi-modal imaging system is adapted to obtain three modalities of the tumor.

[0067] The first modality is positron emission tomography.

[0068] Such modality is often named by the acronym PET and designates a technique that measures physiological function by looking at blood flow, metabolism, neurotransmitters, and radiolabelled drugs. PET offers quantitative analyses, allowing relative changes over time to be monitored as a disease process evolves or in response to a specific stimulus.

[0069] The first modality provides with a metabolic view of the tumor.

[0070] A metabolic view is defined as a medical image reporting on a metabolic parameter, such as the rate of glucose utilization, which is a favored parameter for most tumors because of the capacity of metabolic views to detect the Warburg effect, a typical feature of cancer. However, the method for predicting may also apply to other metabolic views, such as the presence of a receptor, transporter, enzyme or any molecular biomarker specific of the tumor.

[0071] The second modality is ultrasonic imaging.

[0072] More specifically, the second modality is ultrafast ultrasonic Doppler.

[0073] Such second modality is often named by the acronym UUDI.

[0074] The second modality provides a vascular view of the tumor.

[0075] The third modality is computed tomography.

[0076] Such third modality is often named by the acronym CT and designates a medical imaging technique used to obtain detailed internal images of the body.

[0077] More specifically, the third modality is here a X-ray computed tomography.

[0078] The third modality provides with morphological features of an area.

[0079] An example of such multi-modal imaging system is a PETRUS system.

[0080] PETRUS stands for “Positron Emission Tomography Registered Ultrafast Sonography”.

[0081] The PETRUS system allows for simultaneous acquisition of tissue metabolism using 2’-[18F]fluoro-2’deoxy-D-glucose (FDG) PET, CT and UUDI. The PETRUS system simultaneously reads the cellular metabolism activity alongside the micro-vascular architecture within the tumor, ensuring exact identical physiological conditions during the recording of both sets of spatially co-registered features.

[0082] Images from each modality are taken during the same time interval, corresponding to the duration necessary to acquire images from the slowest of both modalities, so as to ensure that the information acquired from one modality is synchronous in time with the information acquired with another modality.

[0083] Thus, fusion of the images from all modalities is precise, and all the parameters of the tumor are exactly superimposable in the same time interval.

[0084] During the step of extracting, the system 10 extracts a set of parameters from the provided images.

[0085] The set of parameters is a specific set of parameters.

[0086] The set of parameters comprises at least three parameters P1 , P2 and P3.

[0087] The first parameter P1 is the volume of the tumor calculated based on images obtained by one modality.

[0088] The modality is chosen among ultrasound imaging and positron emission tomography and computed tomography imaging, according to the available images. Advantageously, several first parameters P1 are provided, so that the first parameters P1 comprise several values for volume of the tumor, each value being obtained by calculation on images obtained by a respective modality.

[0089] According to a first example, the first parameters P1 are a tumor volume value for the ultrasound imaging modality and a tumor volume value for the positron emission tomography modality.

[0090] According to another example, the first parameters P1 are three tumor volume values, a first tumor volume value for the ultrasound imaging modality, a second tumor volume value for the positron emission tomography modality and a third tumor volume value for the computed tomography imaging modality.

[0091] In the present case, the PETRUS system provides with images according to the three modalities.

[0092] The second parameter P2 is a parameter relative to the interaction between a chemical agent and the tumor.

[0093] For instance, the second parameter P2 is the concentration of a chemical agent in a volume containing the tumor.

[0094] Other second parameters P2 may be considered such as those that define the binding parameters of the chemical agent, or the uptake of the agent by the tumor, including but not limited to, cellular uptake, retention, substrate utilization, binding potential, relative specific tissue uptake, metabolic rate of glucose utilization, and, in general, any parameter that has been demonstrated to quantify the molecular composition or activity of a tumor.

[0095] The chemical agent is chosen among drugs targeting a specific aspect of the tumor. In a specific embodiment, the chemical agent is inhibitor of the vascular endothelial growth factor receptor VEGFR.

[0096] The third parameter P3 is a parameter relative to the presence of tumor vessels in a volume containing the tumor.

[0097] For instance, the third parameter P3 is the number of nodes, a node corresponding to a bifurcation of a vessel.

[0098] Another example of third parameter P3 is the density of nodes.

[0099] According to still another example, the third parameter P3 is the length of at least one vessel in a volume containing the tumor.

[0100] For instance, the third parameter P3 is chosen among the average of the minimum length of each vessel, the average of the minimum length of each vessel, and the variance of the mean vessel lengths. The extracted set of parameters may comprise other kind of parameters, such as the tortuosity of vessels, the density of vessels or the entropy of glucose consumption.

[0101] Preferably, the extracted set of parameters consists in a combination of one or several first parameter(s) P1 , one or several second parameter(s) P2, and one or several third parameter(s) P3.

[0102] During the step of determining, the system 10 determines the tumor stage based on the extracted set of parameters.

[0103] Each tumor stage is associated to a domain of values for each parameter of the set of parameters.

[0104] Thus, the system 10 determines the tumor stage based on the domain to which the extracted set of parameters belongs.

[0105] More specifically, the system 10 first determines the domain to which the extracted set of parameters.

[0106] Then, the system 10 reads in a memory the tumor stage associated to the determined domain.

[0107] This tumor stage that is read by the system 10 is the tumor stage to be determined.

[0108] In the present example, the tumor number of stages is equal to 4.

[0109] The first stage corresponds to an early stage of the tumor.

[0110] The second stage corresponds to a first intermediate reversible stage of the tumor.

[0111] The third stage corresponds to a second intermediate reversible stage of the tumor.

[0112] The fourth stage corresponds to an irreversible end stage of the tumor.

[0113] The stages are defined by specific ranges of values for each parameter. In a specific embodiment, the mean values of each parameter along with their corresponding standard deviations are provided, as shown in Table 3.

[0114] More generally, the tumor stages may be obtained by implementing a specific computer-implemented process.

[0115] In such process, images of the tumor of many subjects are provided to a calculator.

[0116] The calculator applies an extraction technique on the provided images of the many subjects. This enables the calculator to obtain extracted parameters.

[0117] Such extracted parameters comprise at least one metabolic parameter, at least one vascular parameter and at least a morphological parameter.

[0118] Then, the calculator eliminates the redundant parameters from the extracted parameters by applying a pretreatment on the extracted parameters. This enables the calculator to obtain selected parameters. The calculator applies a Hierarchical Clustering Analysis on the selected parameters to obtain the tumor stages and the associated domain.

[0119] Such process is schematically illustrated by figure 3, which will be described in reference to the experiments carried out by the Applicant in the next section.

[0120] These experiments show that the method for determining enables to provide with a view of the response of paraganglioma tumors to antiangiogenic therapy.

[0121] Indeed, this approach discriminates different stages corresponding to the response or escape from treatment with the anti-angiogenic drug sunitinib, and defining the individual trajectories of response to treatment. In contrast to RECIST or PERCIST criteria, the method documents stages of tumor progression under anti-angiogenic treatment, and could be generalized to other precision drugs targeting functional and molecular hallmarks of cancer.

[0122] This response is provided with only multimodal imaging of the tumor, which renders such determination easy to implement.

[0123] The method can thus help to better understand precision anti-cancer therapies by providing image-based insight about individual trajectories. The approach could also set a basis for personalized follow-up of cancer treatment with targeted therapies.

[0124] It is also interesting to note that the present method can be used in other type of treatments or cancers.

[0125] In particular, the method can be applied to chemotherapy, radiation therapy, hormonal therapy or immune therapy.

[0126] EXPERIMENTAL SECTION

[0127] MATERIALS AND METHODS

[0128] Figure 3 shows the pipeline of the study, from the acquisitions of PETRUS image volumes to the definition of individual trajectories of response to treatment. Hierarchical clustering and Machine learning classifiers were built using MATLAB version R2021 -b.

[0129] More specifically, figure 3 illustrates that simultaneous co-registered PETRUS image volumes were processed to extract features describing the metabolic, vascular and morphological components of tumor development. A Pearson correlation study was performed to remove redundant features. Features describing all the time points were combined and Hierarchical Clustering Analysis was applied to obtain clusters representing different states of the tumor. The clusters and classes identified with HCA were used with 10 different supervised machine-learning classifiers for model generalization and final validation. Finally, time-wise concatenation of the identified stages was performed to form the individual trajectories of tumor evolution for each animal.

[0130] Acquisition of live animal imaging data

[0131] Animal experiments were approved by the French Ethical committee under reference No 16-098 and performed by certified personal following the French law on animal experimentation No 2013-1 18.

[0132] Two groups of mice followed the protocol of tumor implantation, sunitinib treatment, imaging, and physiological follow-up described in an article named Facchin et al. 2020 and entitled “Concurrent imaging of vascularization and metabolism in a mouse model of paraganglioma under anti-angiogenic treatement” (Facchin et al., 2020 hereinafter). The first group (training group) included 16 mice from (Facchin et al., 2020), while the second group (validation group) included another 1 1 mice that underwent the same experimental protocol. Imaging of the training group was performed immediately before the beginning of treatment (baseline), and then every week until week 3 for vehicle-treated animals (8 mice), and every week until week 6 for sunitinib-treated animals (8 mice). The validation group contained only sunitinib-treated animals and imaging was performed at baseline, week 1 , week 3 and week 6.

[0133] This is schematically represented on figure 4, which describes the database generation process. Mice in the training group were divided into two groups: sunitinib and control (SUNI and CON, respectively). Eight mice from each group were scanned with PETRUS before and after 1 , 2 and 3 weeks of treatment. SUNI mice were further imaged at 4, 5 and 6 weeks of treatment. Mice from the independent validation set were scanned at baseline and weeks 1 , 3 and 6 of sunitinib treatment.

[0134] In brief, adult female nude mice 6-week-old weighing 30g (Janvier Labs, France) were implanted in the dorsal fat pad with tumors obtained from immortalized mouse chromaffin cells (imCC) carrying a homozygous knock-out of the Sdhb gene (Sdhb / ). Mice were housed under controlled temperature (24°C), relative humidity (50%), a 12 / 12 light / -dark cycle, and free access to water and food. Whenthe tumor volume reached 140 mm3, mice were randomly divided into a vehicle group (CON, n = 8) and a sunitinib group (SUNI, n = 8). The sunitinib group received sunitinib malate (Clinisciences, A10880-500) daily at a dose of 50 mg / kg body weight during 6 consecutive weeks, administered by oral gavage of 200 L in a 10 mg / mL DMSO / PBS (1 :4) solution. The control group received the DMSO-PBS solution (1 :4) daily during 3 weeks. Mice were humanely euthanized when the tumor volume reached UKCCCR recommendations, or if they showed signs of advanced cancer disease.

[0135] The effect of sunitinib was monitored non-invasively using the hybrid in vivo imaging technology PETRUS (Positron Emission Tomography Registered Ultrafast Sonography), that allows for simultaneous acquisition of tissue metabolism using 2’-[18F]fluoro-2’deoxy- D-glucose (FDG) PET, Computed Tomography (CT) and Ultrafast Ultrasound Doppler Imaging (UUDI). PETRUS simultaneously reads the cellular metabolism activity alongside the micro-vascular architecture within the tumor, ensuring exact identical physiological conditions during the recording of both sets of spatially co-registered features.

[0136] Description of Database formation

[0137] Each PETRUS acquisition comprised three image volumes registered in a common time and space reference frame that defined a multiparametric cube including the whole tumor. The features describing the metabolic, vascular, and morphological characteristics of the tumor were extracted from the PET, UUDI, and CT images, respectively. A Volume of Interest (VOI) covering the whole tumor was defined on the PET images by segmenting voxels with a FDG Standard Uptake Value (SUV) greater than 30% of the tumor’s peak SUV at 50-60 minutes post-injection. This VOI was used to create a binary mask that was applied to the time and space registered volumes from each imaging modality. From the masked PET image, the following metabolic features were extracted: mean, standard deviation, minimum and maximum of Standard Uptake Values (MeanSUV, StdSUV, MinSUV, MaxSuv), and PET volume (PETVolume).

[0138] The masked UUDI volume was filtered using a Hessian-based vessel enhancement filter, and vessels were segmented using predefined thresholds and skeletonized using an iterative ordered thinning-based skeletonization method. The skeletonized mask of vessels was transformed into a graph of nodes and edges representing the vascular network of the tumor. Using this graph, the following features describing the topology of the tumor vascularization were calculated: mean, minimum and maximum vessel length (MeanVesselsLength, MinVesselsLength, MaxVesselsLength), mean vessel tortuosity (Tort), where the tortuosity of one vessel is calculated by dividing the length of the vessel between two nodes divided by the shortest distance between the nodes, vessels length dispersion (VesselsLength-Disp), which is the coefficient of variation (standard deviation divided by the mean) of the vessels lengths, number of nodes (NumNodes), density of nodes (DensityNodesinUSV), mean vessels diameter (MeanVesselsDiam) and ultrasound volume (USVolume), which is the number of voxels of the vascular skeleton multiplied by the voxel volume.

[0139] A more detailed description of the extracted features is presented in Table. 1. The quantification of PETRUS images was performed using MATLAB® version R2021 b. The CT volume (CTVolume) was manually delineated from the fat pad surrounding the tumor.

[0140] The working database assembled all 17 features extracted from the imaging modalities as a unique record number that defined the mouse the week of the imaging session (week zero (WO) being the pre-treatment imaging session), and the treatment group assignment (CON for sham-treated mice, SUNI for sunitinib-treated mice). Data were divided into 3 subgroups,

[0141] (i) D^ningcontaining the SUNI mice in the training group (54 records),

[0142] (ii) Deining containing the CON mice from the training group (27 records), and

[0143] (iii) D^aitdat containing the SUNI mice from the validation group (28 records).

[0144] Feature selection

[0145] Feature selection is an important pre-processing step that affects the learning rate and quality of any classifier. Removing non-useful or redundant features reduces the dimensionality of the feature space, which is an essential step to improve the performance of the classifier (Guyon and Elisseeff, 2003). In order to identify linear correlations between the different features, we applied a Pearson correlation using a Pearson coefficient r of 0.9 (p-value of 0.05) to detect redundant features.

[0146] In addition, non-informative features with a low coefficient of variation (CV of 0.1 ) were removed.

[0147]

[0148] Table 1: PET / CT / UUDI extracted features. N / A : no unit Unsupervised Classification: Hierarchical Clustering

[0149] Hierarchical Cluster Analysis (HCA), an unsupervised machine-learning clustering approach, was used to stratify common metabolic and vascular phenotypic patterns of PGL tumors into clusters, each cluster defining the state of a tumor under sunitinib or sham treatment as a representative combination of metabolic, morphological and vascular features. All individual records were assigned to a cluster, independently of the treatment duration. HCA was applied separately on the Diningar|d Dtratning data-sets from which the treatment time points (Wn) had been removed. Each of the training datasets was then standardized separately by z-scoring, using the mean and the standard deviation of each dataset. The proximity of individual records was measured by computing the unweighted average Euclidean distance. The average link was then computed as a similarity metric to define the closest pair of clusters. Finally, clusters were displayed in a heat map with dendrograms, the length of the dendrogram branches connecting the records or the features being inversely proportional to the similarity of their profiles. Gap statistics (Tibshirani et al., 2001 ) was applied in order to evaluate the optimal number of clusters, and Kruskal-Wallis test was applied to identify significantly different clusters (?). The final result of this analysis provided the optimal number of clusters corresponding to a particular phenotype identified for each data-base record. HCA and statistical tests were implemented in MATLAB® (version 2021 -b) using the clustergram, ttest2, and evalclusters functions, respectively.

[0150] Supervised Classification: Model Building and Validation

[0151] To test the stability of the method, the Applicant compared the clustering results applied on an external population ( D^dlat) to a classification produced as a generalization of the clustering performed on the Applicant’s initial population (Draining) - More precisely, the Applicant considered the clusters of the initial population (P ^ning) as classes of a supervised classification algorithm to predict the classes expected in the new population ^Dvaiidat) - Because the Applicant’s training dataset has an unbalanced number of instances per class, which can undermine the predictability of the models, we performed oversampling through the synthetic minority over-sampling technique (SMOTE), which balances the minority classes. This technique uses the k-nearest neighbors approach to synthesize new observations based on the existing records, the Applicant applied smote using the four nearest neighbors to balance each of the four clusters (A, B1 , B2, and C).

[0152] The selected features of our D “ ingWere brought into ten machine learning classifiers, including decision tree (DT), Gaussian naive Bayes (GNB), kernel naive Bayes (KNB), linear support vector machine (Linear SVM), quadratic support vector machine (Quadratic SVM), k-nearest neighbors (KNN), weighted k-nearest neighbors (Weighted KNN), random forest (RF), narrow neural network (Narrow NN), bilayered neural network (Bilayered NN). The best-performing model was selected by comparing the area under the receiver operating characteristic curve (AUG) and accuracy (ACC) values. The control parameters of the best model were further optimized by Bayesian optimization and five-fold cross-validation to evaluate the performance of the classifier. All classifiers were trained and validated using the classification learner application implemented in MATLAB version 2021 -b. In order to check the relative importance of each of the metabolic, vascular, and morphological features in the classification problem, the Applicant used the predictor importance attribute associated with the RF model. The predictor importance attribute is an implicit technique performed using the RF model and is evaluated using the Gini impurity criterion index. This index is based on the principle of impurity reduction to provide the power of each feature in the classification.

[0153] Identification of Trajectories of Treatment Responses

[0154] The Applicant then tested whether the records assembled within each cluster, corresponding to a tumor state with specific biomarkers, could represent a chronological stage of tumor evolution. By referring back to the time point of each record (week after treatment initial date) in both the CON and SUNI groups, the clusters were ordered chronologically and a time-dependent trajectory was obtained for each mouse. The Applicant applied an R2 test on the 7 timepoints and the treatment response states (clusters obtained from the HCA) to check if these states reflected chronological treatment response stages. Finally, the transitional matrix between clusters was analyzed.

[0155] RESULTS

[0156] Pearson Correlation

[0157] In order to eliminate redundant features, a Pearson significance of |r| > 0.9 and a p- value < 0.05 were applied to all pairs of the 17 features for the 7 timepoints of the Dining- This reduced the number of vascular features from eleven to eight. MeanVesselsLength was highly correlated with MeanVesselsDiameter, Tort, and VesselsLengthDisp; VesselsLengthDisp highly correlated with MeanVesselsDiameter and Tort, and Tort highly correlated with MeanVesselsDiameter. One representive parameter was randomly selected from the sets of highly-correlated parameters. Thus, MeanVesselsLength, MeanVesselsDiameter and Tort were not considered individually but were represented by VesselsLengthDisp. Applying the same Pearson r and p values reduced the metabolic features from five to four. MeanSuv correlated with MaxSuv, and MaxSuv was not considered further. Interestingly, with respect to vascular-metabolic correlations, CVStdSUV significantly correlated with MeanVessels-Diam.

[0158] In addition, since a low coefficient of variation (CV < 0.1 ) results is non-informative for the training of a classifier, features with high Pearson correlation and low coefficient of variation were not considered further. Overall, 8 features including 4 vascular features, i.e. USVolume, NumNodes, DensityNodesinUSV, VesselsLengthDisp, 3 metabolic features, i.e. CVStdSUV, PETVolume, MeanSUV, plus the CT volume, were applied to the three curated databases

[0159] Hierarchical Clustering approach

[0160] Sham-treated Training set (D^ning)

[0161] Performing the hierarchical clustering on thetc°^nfnpdataset identified two major clusters: Clusters Acand Cc, where subscript “c” stands for the control group. Typical feature values are reported in Table 2. Table 2: Metabolic, vascular and morphological characteristics of the clusters from the Dl°iningdataset, presenting the average value for each parameter of each cluster. Mean value in black, standard error of the mean in parenthesis, mean z-score in blue. Cluster Acwas characterized by low volumes of CT, PET, and UUDI, a high coefficient of variation of the standard deviation of SUV, low number of nodes and low density of nodes. This corresponds to a small-size tumor, with low vascularization and metabolism, and a heterogeneous distribution of FDG uptake.

[0162] Cluster Cc was characterized by high volumes of CT, PET, and UUDI, a significantly lower coefficient of variation of the standard deviation of SUV, and a high number of nodes.

[0163] This corresponds to a tumor grown to a large volume, with high metabolism and vascularization and low heterogeneity in the distribution of FDG uptake.

[0164] Suniti nib-treated Training set (Deining) The same clustering approach applied to the Dining dataset identified 3 major clusters : Clusters At, Bt, and Ct, where subscript “t” stands for the sunitinib-treated group.

[0165] Moreover, cluster Btseparated into two subgroups: B1tand B2tas appears from table 3:

[0166] Table 3: Metabolic, vascular and morphological characteristics of the clusters from the D training dataset, presenting the mean values for each parameter of each cluster. Mean value in black, standard error of the mean in parenthesis, mean z-score.

[0167] Cluster Atwas characterized by low volumes of CT, PET, and UUDI, a high coefficient of variation of the standard deviation of SUV, and low vessels length dispersion, Number of nodes and density of nodes. This corresponds to a small-sized tumor with low vascularization and heterogeneous distribution of FDG uptake value, features that are similar to those of cluster Acof the control group.

[0168] Cluster Ctwas characterized by high volumes of CT, PET, and UUDI, low coefficient of variation of the standard deviation of SUV, high vessels length dispersion and very high Number of nodes. This cluster corresponds to a tumor with a large volume, high metabolism and vascularization, and low heterogeneity in the distribution of FDG uptake, features that are similar to those of cluster Ccof the control group.

[0169] To compare the A and C clusters obtained with the SUNI and CON groups, respectively, a Kruskal-Wallis test (Kruskal and Wallis, 1952) was performed between the At and Acclusters, and also between the Ctand Ccclusters. The clusters were statistically similar (p-value j 0.05), indicating that clusters Atand Acon the one hand, and clusters Ctand Cc on the other hand, correspond to similar tumor states in the sunitinib treated and sham treated groups.

[0170] In the SUNI group, the HCA algorithm identified two further clusters not present in the CON group:

[0171] - Cluster B1twas characterized by low to moderate volumes of CT, PET, and UUDI, low coefficient of variation of the standard deviation of the SUV, high vessels length dispersion and a very high density of nodes. This corresponds to a small tumor with a significant but moderate level of vascularization, and medium-to-high heterogeneity in the distribution of FDG uptake.

[0172] - Cluster B2twas characterized by moderate volumes of CT and PET, high UUDI volume, lower coefficients of variation of the standard deviation of SUV, high vessels length dispersion and low density of nodes. This corresponds to a moderate to high tumor volume and vascularization, and low heterogeneity in the distribution of FDG uptake.

[0173] Robustness of Clusterization

[0174] An additional validation step was performed in order to ascertain that clusterization formation was reproducible and not a casuistic process. HCA clustering was repeated on subsets of random instances of the group, formed by randomly removing one mouse at a time. The accuracy of each HCA was calculated by considering the clusters obtained for all mice as ground truth, and comparing it with the clusters of the new subset using the following formula:

[0175] Number of correct predictions

[0176] Accuracy = - - - - - - - - -

[0177] Total number of predictions

[0178] As shown in Table 4, accuracy for every HCA was greater than 95% for the 3 major clusters (At, Bt, Ct).

[0179] Table 4: Performance of HCA for subsets of the Draining dataset. Each of the eight data subsets correspond to the removal of all time points of one mouse from the complete dataset.

[0180] Performance of Supervised Machine Learning Models

[0181] All 10 of the ML classifiers explored demonstrated good predictive performance. GNB achieved the best predictive performance (AUG: 100, ACC: 98.7), whereas DT exhibited the weakest (AUG: 96, ACC: 94.8). The remaining classifiers achieved the following predicted performance: Quadratic SVM (AUG: 100, ACC: 97.4), KNB (AUG: 98, ACC: 94.8), Linear SVM (AUG: 100, ACC: 97.4), KNN (AUG: 97, ACC: 98.7), RF (AUG: 100, ACC: 94.8), Narrow NN (AUG: 100, ACC: 96.1 ), Bilayered NN (AUG: 98, ACC: 94.8) and Weighted NN (AUG: 100, ACC: 97.4).

[0182] Applying the best classifier to the three records that had not been classified using HCA, i.e., mouse 1 -week 6, mouse 3-week 6, and mouse 8-week 5, allowed to classify these records into clusters Ct , Ct , and At , respectively (Table 5).

[0183]

[0184] Table 5: Evolutionary path of sunitinib-treated mice of training set. Items marked as indicate the absence of corresponding PETRUS data. Clusters that were assigned using the RF model are underlined.

[0185] This classification remained consistent with the previous stages of the sunitinib training set The best trained model applied to the D^dlatdataset assigned a state for each record and mouse that was consistent with the states of the (Draining) dataset, as appears from table 6:

[0186] Finally, the feature importance attribute of the Random Forest, showed that all three types of tumor features, i.e., metabolic, vascular and morphological features, participated in the prediction of the four clusters. This indicates that the information provided by each of the 3 imaging modalities all contributed in defining tumor stages for each imaging record.

[0187] Clusterization reveals tumor progression

[0188] The Applicant then tested whether the different clusters would correspond to different timepoints during the tumor follow-up, i.e., whether, for any record, there was a correlation between assignment to one particular cluster and the timepoint at which imaging had been performed for that record. Regarding the CON group, all except two records (mouse 3 / week 2 and mouse 6 / week 2) of cluster Ac corresponded to baseline or to the week 1 timepoint. Conversely, all cluster Cc records corresponded to week 2 or week 3 acquisitions. This confirms that cluster Acrepresents an initial stage of the tumor, while cluster Ccrepresents an advanced tumor stage.

[0189] In contrast, the correspondence between the timepoint of acquisition and assignment to the At or Ct cluster was much looser for the SUNI group than for the CON group. For example, mouse 6 remained in cluster At at all time points until week 6. Moreover, at baseline and week 1 , a significant number of mice were not assigned to the Atcluster, but either to the B1tcluster (2 mice at baseline and 3 at week 1 ) or to the B2tcluster (two at each timepoint). Conversely, upon reaching the last observation timepoint (week 6), 5 mice from the SUNI group were in the Ctcluster, while one was classified in the B1tcluster, one in the B2tcluster, and one in the Atcluster.

[0190] The Applicant then asked what was the influence of the vascular and metabolic features on the clustering results. Removing PET and UUDI features from the SUNI datasets and basing clustering only on CT volume led to co-clustering of [At;B1t] and [B1t;B2t]. This indicates that RECIST-like criteria using CT only, did not identify intermediate clusters. When the same algorithm HCA was applied to the SUNI dataset from which the vascular features obtained by ultrasound imaging had been removed, i.e., using only the PET metabolic features and the CT volume, only 2 significantly different clusters were obtained using gap statistics: clusters APET / CT and BPET / CT. This, indicates that PERCIST -like criteria, using PET-CT only, did not identify intermediate clusters. Therefore, the intermediate B stage (Bt) and its two sub-clusters B1tand B2t, reflect essentially changes concerning the vascular features of tumors under sunitinib treatment.

[0191] Clusters depict responses to sunitinib treatment

[0192] To further understand how clusters reflect the response to sunitinib treatment, the evolutionary trajectories (passage from one cluster to another over successive timepoints) were studied individually for each mouse of the SUNI group (Table 5).

[0193] The progression from cluster Atto Ctof sunitinib-treated mice was not direct as the Acto Ccin sham-treated animal, but passed through intermediate Btclusters. This was confirmed by a correlation analysis performed on clusters At, Bt, Ctconsidered as stages 1 , 2 and 3, resulting in R2= 0.84. Calculation of the cluster transition matrix confirmed the relationship between the clusters and the chronology of tumor evolution. Assuming a progression represented by states At, Bt, and finally Ct, the Applicant obtained 29 / 46 (65.9%) stable phenotypes, i.e., remaining in the same state; 10 / 46 (22.7%) one progression, i.e., advancing further to the next state, and 5 (11 .3%) regressions from Btto At (Table 7-Supplementary). Pooling the validation population and the training population showed a dissymmetry between ’’progression” (n=15) and ’’regression” (n=6). Finally, cluster Ctwas an irreversible transition deriving essentially from the B2tstate, that appeared as an obligatory intermediate stage to reach state Ct, and the transition from Btto Atoccurred only by the intermediary of the stage B1t, not B2t. States At, B1t, B2tand Ctare thus ordered in time, suggesting that they are in fact tumor stages and that there is a progressive evolution of tumor stages from states Atto Ctthrough Bt, and irreversibility between Ctand the other states.

[0194] In summary, multi-feature ML analysis of the sunitinib-treated animals showed that individual trajectories, defined by passage from one cluster to another, followed a discrete number of rules:

[0195] - Irrespective of whether mice received sunitinib or vehicle, no mouse reversed from the advanced tumor stage (cluster C) to a less advanced stage.

[0196] - In the sunitinib group, mice moved from the early tumor stage (cluster At) to either one of the two intermediate stages (clusters B1tor B2t), but not directly to the advanced stage (cluster Ct). - In the sunitinib group, mice moved from cluster B11to B2tand back, and from cluster B1tback to cluster At, but no passage from cluster B2tto cluster Atwas observed.

[0197] - In the sunitinib group, all mice reaching the advanced (cluster Ct) stage originated from cluster B2t.

[0198] The robust correlations between clusters and treatment duration, and the transition matrix between clusters confirm that the A, B and C clusters correspond to tumor stages. Interestingly, transitions between sub clusters B1tand B2twere less correlated with time than transitions between Atand B1tor B2t, and between B2tand Ct. This suggests that the ’’reverse” transitions, i.e., B2tto B1tand B1tto At, could reflect the phenotypic changes associated with a positive response to sunitinib.

[0199] Figure 5 summarizes the trajectories between tumor stages in sunitinib-treated mice. There was first an increase in the level of tumor vascularization (At-> B1ttransformation), followed by a decrease in the heterogeneity of FDG distribution in the tumor (B1t-> B2t).

[0200] Discussion

[0201] Previous studies have used Machine learning (ML) to study the correspondence between gene expression and tumor progression. This is the first application of ML based on Hierarchical Clustering (HCA) and supervised ML algorithms to noninvasive imaging of PGL. PGL lesions may concern the whole sympathetic and parasympathetic chains from the base of the skull to the pelvis. Germline mutations in one of the SDHx genes are responsible for approximately 20% of cases of PGL and also in some other tumors. PGL patients carrying SDHx mutations show a higher rate of metastatic disease and a lower rate of survival than non-SDHx PGL patients. Surgery is not without risk and may be impractical for numerous or misplaced lesions, and clinical trials with sunitinib have reported modest results in SDHB mutation carriers.

[0202] There is an international consensus on the use of repeated noninvasive imaging for the screening, management and follow-up of PGL patients, as well as for asymptomatic SDHx mutation carriers. The Applicant’s results show that unsupervised machine learning of serial noninvasive imaging data can define the phenotypic stages of mouse Sdhb- / - PGL tumors under anti-angiogenic treatment. The main finding is that, although the records fed to the ML algorithm had not been time-stamped for the duration of treatment, unsupervised ML applied to multimodal multiparametric imaging features yielded clusters relevant to disease progression and to the response to sunitinib. In the sham-treated group, all mice switched, generally in less than three weeks, from cluster Ac, an early stage with small and poorly developed tumors, low vascularization and heterogeneous FDG uptake, to cluster Cc, an advanced stage with large tumors, large vessels, high and relatively homogeneous FDG uptake, corresponding to an end-stage cancer disease. In the sunitinib-treated group, a given tumor from a given mouse could, over time, move from one cluster to another, suggesting that the changes from one cluster to another depicted trajectories of tumor evolution related to the response or the escape from treatment. Some sunitinib-treated tumors showed a progression similar to sham-treated tumors, which infers that sunitinib- treated mouse entering the advanced-stage Ctcluster have escaped sunitinib treatment.

[0203] Two other clusters, B1tand B2t, representing intermediate tumor stages, were observed only in the sunitinib-treated group, supporting the view that their phenotypes represent the effects of sunitinib on PGL tumors. The first one, B1t, encompassed smallsized tumors with a significant but moderate level of vascularization and heterogeneity in the distribution of glucose uptake. The second cluster, B2t, encompassed tumors of moderate volume and vascularization, and low heterogeneity in the distribution of glucose uptake. ML did not identify these two intermediate stages when the vascular features derived from ultrafast ultrasound were removed from the analysis. Therefore, the B1tand B2tintermediate stages identified the effect of sunitinib on tumor vascularization, likely by inhibition of vascular endothelial growth factors receptors (VEGFRs), the major pharmacological target of the drug.

[0204] Interestingly, all of the three mice that pertained to a B cluster (B1tor B2t) at baseline ended up in the Ctcluster at the end of the 6-week sunitinib treatment, while only one of the 4 mice pertaining to the Atcluster at baseline ended up in the Ctcluster. Although further studies are necessary to determine whether the tumor’s biology prior to the administration of sunitinib could predict future escape from treatment, this may indicate that tumors that have already developed a significant vessel network are less prone to respond to sunitinib therapy. Thus, even though the switch from B1tto B2twas reversible under sunitinib treatment (B1tB2t), increased vascularization and decreased metabolic heterogeneity defining the B2tstage were necessary features for passage to the Ctstage, in other word for escape from sunitinib treatment. From a cancer biology point-of-view, this suggests that escape from sunitinib treatment involves both a metabolic and a vascular switch.

[0205] Moreover, clusters’ extraction using multiple features could allow to gain a better understanding of the sequence of events underlying drug response. The fact that cancer is a multiform disease with multiple intermingled hallmarks has been extensively documented,. Therefore, it is unlikely that assessing only one biomarker, even one that informs on the activity towards the pharmacological target, may be sufficient to assess treatment response, and, even less so, to identify complex escape mechanisms. All in all, the present results support the recourse to multimodal imaging with careful selection of relevant imaging biomarkers, ideally including one or several biomarker(s) of the hallmark targeted by the treatment. In this respect, other tumor variants could also benefit from similar approaches extracting biomarkers specific for the tumor type and / or treatment. Finally, it may also be interesting to apply a radiomic analysis in order to compile mathematically- defined image features, and determine whether they represent phenotypic states predictive of tumor stage predictive of treatment response.

[0206] Conclusion

[0207] The combination of hierarchical clustering and supervised machine learning algorithms provides remarkable insight into the progression of tumor development in a mouse model of paraganglioma. Through the incorporation of multi-modal information, including the vascular features of the tumor-targeted by sunitinib, our approach is successful in depicting trajectories of response to treatment. This approach could set a basis for personalized follow-up of tumors treated by targeted therapies.

Claims

CLAIMS1 A method for predicting the tumor stage of a tumor of a subject, the method being computer-implemented, the method comprising the steps of :- providing images of the tumor of the subject obtained by several modalities,- extracting a set of parameters from the provided images, the set of parameters comprising:- a volume of the tumor calculated based on images obtained by one modality,- at least one parameter relative to the interaction between a chemical agent and the tumor, and- at least one parameter relative to the presence of tumor vessels in a volume containing the tumor, and- determining the tumor stage based on the extracted set of parameters.2.- The method for predicting according to claim 1 , wherein the extracted set of parameters comprises several values for volume of the tumor, each value being obtained by calculation on images obtained by a respective modality.3.- The method for predicting according to claim 2, wherein the several modalities at the step of providing comprises ultrasound imaging and positron emission tomography.4.- The method for predicting according to claim 2, wherein the several modalities at the step of providing consists of positron emission tomography, ultrasound imaging and computed tomography imaging.5.- The method for predicting according to any one of the claims 1 to 4, wherein the at least one parameter relative to the interaction between a chemical agent and the tumor is the concentration of the chemical agent in a volume containing the tumor.6.- The method for predicting according to any one of the claims 1 to 5, wherein the at least one parameter relative to the presence of tumor vessels is the number of nodes, the density of nodes or a length of at least one vessel.7.- The method for predicting according any one of the claims 1 to 6, wherein the at least one parameter relative to the length is chosen in the list consisting of :- the average of the minimum length of each vessel,- the average of the minimum length of each vessel, and- the variance of the mean vessel lengths.8.- The method for predicting according to any one of the claims 1 to 7, wherein the set of parameters consists in:- a volume of the tumor calculated based on images obtained by several modalities,- one parameter relative to the concentration of a chemical agent in a volume containing the tumor, and- one parameter relative to the presence of tumor vessels in a volume containing the tumor.9.- The method for predicting according to any one of the claims 1 to 8, wherein each stage is associated to a domain of values for each parameter and, during the step of determining, the stage is determined based on the domain to which the extracted set of parameters belongs.10.- The method for predicting according to any one of the claims 1 to 9, wherein the number of stages is equal to 4 and wherein:- the first stage corresponds to an early stage of the tumor,- the second stage corresponds to a first intermediate reversible stage of the tumor,- the third stage corresponds to a second intermediate reversible stage of the tumor, and- the fourth stage corresponds to an irreversible end stage of the tumor.11 .- The method for predicting according to any one of the claims 1 to 10, wherein the tumor is a paraganglioma.12.- The method for predicting according to any one of the claims 1 to 11 , wherein the subject is a mammal.13.- A computer program product comprising instructions for carrying out the steps of a method according to any one of claims 1 to 12 when said computer program product is executed on a suitable computer device. 14.- A computer readable medium having encoded thereon a computer program product according to claim 13.

Images