Method for classifying a lesion as a uterine leiomyoma or as a uterine leiomyosarcoma

Abstract

A computer-implemented method (100) for classifying a lesion as a uterine leiomyoma, UM, or as a uterine leiomyosarcoma, LMS, the method (100) comprising: acquiring (101) at least one Computed Tomography, CT, image of the lesion; obtaining (103) a Region of Interest, ROI, of the lesion within the at least one CT image; calculating (105) one or more radiomic features of said ROI of the lesion; selecting (107), using a plurality of feature se-lection functions, a corresponding plurality of sub-sets of features, wherein each of said plurality of subsets of features comprises one or more selected radiomic features of said one or more radiomic features; applying (109) a corresponding plurality of predictive models which have been trained for classifying a lesion as UM or as LMS to the plurality of subsets of features; and classifying (113) the lesion as UM or as LMS using said plurality of the predictive models.

Description

Method for classifying a lesion as a uterine leiomyoma or as a uterine leiomyosarcomaDescriptionTECHNICAL FIELD

[0001] The present disclosure concerns a computer-implemented method for classifying a lesion as a uterine leiomyoma or as a uterine leiomyosarcoma, which can accurately discriminate between uterine leiomyoma or uterine leiomyosarcoma.

[0002] The present disclosure also concerns a computer implemented method of training a system for classifying a lesion as a uterine leiomyoma or as a uterine leiomyosarcoma.

[0003] The subject matter disclosed herein also refers to a system, a computer program product and a computer-readable storage medium for performing classification and / or training.BACKGROUND ART

[0004] Uterine leiomyosarcomas (LMS) are rare tumours arising from the muscular uterine wall, which represent 3-7% of uterine malignancies and around 1 % of all female genital tract cancers

[0001] , Compared with other types of uterine cancers, they are aggressive tumours with a high risk of recurrence and death, regardless of the stage of the disease at diagnosis [1], Surgical treatment in the early stage consists of total removal of the uterus (hysterectomy) by laparotomy to avoid the high risk of neoplastic spread through rupture and fragmentation of the tumour [2],

[0005] On the other hand, uterine leiomyoma (UM) represents a benign pathology also arising from the muscular wall of the uterus, but very frequent, with an incidence of about 70%-80% [3],

[0006] Discrimination between benign and malignant myometrial lesions is clinically important for planning optimal management (hysterectomy in LMSs, fertility-sparing surgery, medical treatment, or no treatment in UMs) and defining the most appropriate and personalized surgical approach (laparotomy in LMSs versus minimally invasive surgery in UMs) [3,4],

[0007] The accurate and pre-operative diagnosis of LMS is important because of thegrowing availability of more conservative approaches to managing benign uterine masses [5], In contrast, a misdiagnosis can severely impact patient's prognosis, as a morcellated LMSs increases the risk of dissemination of neoplastic material within the abdomen and the possibility of creating an iatrogenic abdominal sarcomatosis [6-8],

[0008] Unfortunately, the diagnosis of LMS is always defined post operation with the definitive histological examination.

[0009] Considering the above and the lack of methods for accurately diagnose LMSs and UMs pre-operation, there is an unmet clinical need to provide an accurate method for correctly diagnose LMSs and UMs without the necessity of an intraoperative procedure, which is invasive and can increase the risk of dissemination of neoplastic material.SUMMARYThe present application provides a method designed to classify a lesion as LMS or as UM, using computer tomography images. This method achieves optimal level of accuracy in the differential diagnosis of these tumours.The present application also provides a computer implemented method of training a system for classifying a lesion as a uterine leiomyoma or as a uterine leiomyosarcoma.The invention therefore covers a classification method as claimed in claim 1 , a training method as claimed in claim 8, a system as claimed in claim 16, a computer program product as claimed in claim 17 and a computer-readable storage medium as claimed in claim 18.Further, preferred advantageous embodiments are described in the dependent claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0010] A more complete appreciation of the disclosed embodiments of the invention and many of the attendant advantages thereof will be readily obtained as the same becomes better understood by reference to the following detailed description when considered in connection with the accompanying drawings, wherein:Figure 1 shows a schematic view of a system for classifying a lesion as a uterine leiomyoma or as a uterine leiomyosarcoma, according to the present disclosure.Figure 2 illustrates a flowchart of a method for classifying a lesion as a uterine leiomyoma or as a uterine leiomyosarcoma, according to the present disclosure.Figure 3 illustrates a flowchart of a method of training a system, such as the system shown in figure 1 , for classifying a lesion as a uterine leiomyoma or as a uterine leiomyosarcoma, according to the present disclosure.DETAILED DESCRIPTION OF EMBODIMENTS

[0011] Reference is made to figure 1 , which shows a schematic view of a system 1 for classifying a lesion, such as a mesenchymal lesion of the uterus, as a uterine leiomyoma (UM) or as a uterine leiomyosarcoma (LMS). The system 1 can also be trained for classifying a lesion as UM or LMS as described with reference to figure 3.

[0012] The system 1 comprises a memory 10 and a processor 20. The system 1 may also comprise a display 30 for displaying the classification output generated by the processing unit 20

[0013] The processing unit 20 is communicatively coupled to the memory 10 and configured to execute a program for performing a method for classifying a lesion as UM or as LMS and / or for training the system 1 .

[0014] The processing unit 20 may comprise one or more processors and a receiving-transmitting module (not shown), coupled with the processor, and configured to receive and transmit the classification output generated by the processing unit 20 to the display 30 and / or a terminal, such as a mobile terminal (MT), a server (S), personal computer or any other remote terminal.

[0015] In some embodiments, the processing unit 20 can be a physical hardware remotely arranged with respect to the memory 10 and / or the display 30. For example, the processing unit 20 may be based or run in a cloud, such as in a terminal or a server that is communicatively coupled via a wired or wireless connection to the memory 10 and / or the display 30.

[0016] The computer program may be stored in a computer-readable storage medium or a dedicated memory of the processing unit 20 (not shown in figure 1 ), and comprise a set of instructions which, when executed by the processing unit 20, cause the processing unit 20 to perform a method for classifying a lesion as UM or as LMS, achieving optimal level of accuracy in the differential diagnosis of these tumours.

[0017] The memory 10 is configured for storage of at least one CT image of a lesion, for example in a database, and can be accessed and controlled by the processing unit 20 in order to retrieve the CT image(s) of lesions and / or store state dataassociated to the method for classifying or training, such as processed CT images, user classification output generated by the processing unit 20, radiomic features, predictive models, etc.

[0018] CT images can include any type of CT image, such as images acquired with different scanners CT1 ... CTn and / or using different acquisition protocols. For example, the CT images can be multiphasic CT images or contrast-enhanced (CE- CT) phase images of the abdomen. CT images can be stored in any format, such as the DICOM™ format.

[0019] Reference is now made to figure 2, which illustrates a flowchart of a method for classifying a lesion as UM or as LMS. This method achieves optimal level of accuracy in the differential diagnosis of these tumours.

[0020] At step 101 the method acquires at least one CT image of the lesion. The image can be acquired from the memory 10 of the system 1 or directly from one or more CT scanners CT1 ... CTn. The acquisition step 101 may comprise acquiring at least one contrast-enhanced computed tomography (CE-CT) image.

[0021] At step 102 the method processes at least one CT image to obtain at least one corresponding processed CT image. Since the CT images may be generated using different scanners CT1 ... CTn and / or protocols, step 102 allows to standardise the CT images allowing for optimal classification. During this step the CT image(s) may be resampled to obtain a pre-set isotropic voxel spacing and a pre-set voxel density. For example, the CT image can be resampled to obtain an isotropic voxel spacing of 5 mm and density discretization using a fixed bin size of 25 Hounsfield Unit (HU).

[0022] The processing step 102 may also comprise adjusting the image by applying one or more filter, including one or more of the following filters: a wavelet filter, a square filter, a square root filter, a logarithm filter, an exponential filter, and a gradient filter.

[0023] At step 103 the method obtains a Region of Interest (ROI) of the lesion within at least one CT image, or the corresponding processed CT image(s) if the images have been pre-processed as provided in step 102 of the method.

[0024] The ROI can be a region drew by an operator, such as a radiologist or agynaecologist, or a region obtained semi-automatically or automatically using a dedicated software, such as, for example, by performing a semantic segmentation process. The ROI can be a two-dimensional region related to a single CT image or a volume of interest (VOI) associated with a group / set of CT images. The ROI / VOI delineates the region of the tumour, such as the region of UM or LMS within the CT image(s), including the tumour tissue that is viable for analysis.

[0025] At step 105 the method calculates / extracts one or more radiomic features of said ROI / VOI of the lesion. The radiomic features can be extracted from each of the CT image(s), or the corresponding processed CT image(s) if the images have been pre-processed.

[0026] The one or more radiomic features of the tumour can include different classes of radiomic features. For example, the extraction / calculation of radiomic features can be performed using a dedicated software, such as the PyRadiomics package in Python™, which offers the possibility to calculate hundreds of features subdivided into 8 different classes: First Order Statistics, 3D Shape-based, 2D Shape-based, Gray Level Cooccurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Size Zone Matrix (GLSZM), Neighbouring Gray Tone Difference Matrix (NGTDM), Gray Level Dependence Matrix (GLDM).

[0027] At step 107 the method proceeds with selecting, using a plurality of feature selection functions, a corresponding plurality of subsets of features. Each subset of features comprises one or more selected radiomic features of the one or more radiomic features calculated / extracted at step 105.

[0028] The plurality of feature selection functions used to perform feature selection can comprise a random forest optimizer function, such as a Boruta™ wrapper (boruta function in R); a regression analysis function, such as a LASSO™ function (glmnet function in R); a recursive feature elimination function, such as an RFE™ function (rfe function in R). The plurality of feature selection functions may comprise any other selection function.

[0029] At step 109 a plurality of predictive models which have been trained for classifying a lesion as UM or as LMS, for example, as described with reference to the training method shown figure 3, is applied to the corresponding plurality of subsets of features obtained at step 107.

[0030] The plurality of predictive models may comprise machine learning-based models that have been trained and tested for the binary classification task based on results of a histological / histopathological examination of UM or LMS, for example by using the method described with reference to figure 3.

[0031] At step 113 the method performs a classification of the lesion as UM or as LMS using the plurality of predictive models.

[0032] The classification step 113 can comprise, for each predictive model, obtaining a probability that the lesion corresponds to UM or LMS based on the predictive model; and classifying the lesion as UM or LMS in response to a determination that said probability reaches a probability threshold value. For example, by using three selection functions is possible to obtain three subsets of features and thus three corresponding predictive models. In this case, the classification step 113 will result in three classification outputs which can be analysed at step 115.

[0033] At step 115 the method performs a cross-correlation analysis of the classification outputs obtained by using said plurality of predictive models; and at step 116 the method determines that the lesion is UM or LMS based on the crosscorrelation analysis. For example, if the majority of classification outputs indicates that the lesion comprises an LMS then the method determines that the lesion is an LMS, or vice versa. Alternatively, if a set number of classification outputs indicates that the lesion comprises an LMS then the method determines that the lesion is an LMS.

[0034] Reference is now made to figure 3, which illustrates a flowchart of a method of training a system 1 for classifying a lesion as UM or as LMS.

[0035] At step 201 the method acquires a set of CT image(s) of the lesions, such as a plurality of images comprising images of various lesions that have been labelled as comprising UM or LMS. These images can be acquired from the memory 10 of the system 1 or directly from one or more CT scanners CT1 ... CTn.

[0036] The method can process the CT images at step 202 to obtain corresponding processed CT images, for allowing optimal training, as detailed in step 102. For example, the CT image(s) may be resampled to obtain a pre-set isotropic voxelspacing and a pre-set voxel density within the images, such as an isotropic voxel spacing of 5 mm and density discretization using a fixed bin size of 25 Hounsfield Unit (HU). In step 202 the method may also adjust the images by applying one or more filters.

[0037] At step 203 the method obtains a ROI / VOI associated with the set of acquired or pre-processed CT images, as detailed in step 103. The ROI / VOI includes all tumour tissue that is viable for performing the training.

[0038] At step 204 the method acquires a set of labels, each being indicative that a corresponding ROI / VOI comprises UM or LMS. The labels can comprise the results of histological / histopathological examinations of the post-surgical specimens of UM or LMS. The set of labels may be stored in the memory 10 of the system 1 , such as in a dedicated database thereof, or may be extracted directly from the set of CT images that are acquired at step 201 .

[0039] At step 205 the method calculates / extracts one or more radiomic features of said ROI / VOI. The radiomic features can be extracted from each of the set of CT images, or the corresponding processed CT images in the event that the images have been pre-processed, as discussed with reference to step 105.

[0040] At step 207 the method proceeds with selecting, using a plurality of feature selection functions, a corresponding plurality of subsets of features. Each subset of features comprises one or more selected radiomic features of the one or more radiomic features calculated / extracted at step 205.

[0041] As per step 107, in step 207 the plurality of feature selection functions used to perform feature selection can comprise a random forest optimizer function, such as a Boruta™ wrapper; a regression analysis function, such as a LASSO™ function; and a recursive feature elimination function, such as an RFE™ function, and / or any other selection function.

[0042] The method proceeds to step 209, in which the method generates a plurality of predictive models that are trained for classifying a lesion as UM or as LMS based on the set of labels.

[0043] The predictive models are generated based on the corresponding plurality of subsets of radiomic features by combining the selected one or more radiomicfeatures of each of the plurality of subsets of radiomic features to fit a corresponding predictive model. The predictive models may comprise generalized linear models (GLMs), which can be fit by executing, for example, the glm function in R.

[0044] The method progresses at step 211 whereby performs a number of cross- validation cycles of the plurality of predictive models to optimize the predictive models generated at step 209, thus increasing the capacity of the predictive models to perform a more reliable classification.

[0045] During the cross-validation cycles, the one or more selected features of each of the plurality of subsets of radiomic features are increased / augmented by applying random seeds to divide the selected features in a training set and a corresponding test set. This operation may be repeated a plurality of times, for example 10-100 times, by splitting the dataset into 70% training and 30% test with balanced output (i.e., the result of a histological / histopathological examination of the post-surgical specimens of UM or LMS that is used to label the CT images). The step of incrementing the plurality of features may also comprise scaling the features using a z-score before performing cross-validation. For each iteration, the features obtained through cross-validation may be stored for performing feature selection.

[0046] The predictive models may be further optimized to obtain a predictive power score for classifying the lesion as UM or as LMS that reaches a predictive power threshold. This may be done by (i) adjusting, for each cross-validation cycle, one or more parameter values that are used to control the learning process of at least one predictive model; (ii) estimating the predictive power score for classifying the lesion as UM or as LMS of the at least one predictive model using the adjusted one or more parameter values; and (iii) determining whether the predictive power score reaches a predictive power threshold.

[0047] For example, an optimal RFE™ model can be obtained after optimization of hyperparameters (i.e. the one or more parameter values that are used to control the learning process) for each cross-validation cycle using, for example, the rfeControl function in R with a 10-fold cross-validation method.

[0048] In another example, the optimal LASSO™ model having the best lambda hyperparameter can be obtained after a 10-fold cross validation using the cv.glmnet function in R.

[0049] Moreover, since the Boruta™ function and RFE™ functions do not consider the collinearity and the correlation between variables, the features selected by these functions can be further reduced with a cross-correlation analysis of the radiomic features obtained by these models, such as by using a Pearson’s correlation test with a Pearson’s r cutoff > 0.60. In this manner, clusters of highly correlated features can be detected and low correlated features from each cluster can be retained.

[0050] At step 213 the method performs an evaluation cycle of the predictive models. The evaluation cycle may comprise, for each predictive model, the steps of: (i) calculating an area, AUC, below a receiver operating characteristic curve, ROC, associated with the predictive model; and (ii) estimating a predictive power score of the optimized predictive model based on said AUC.

[0051] Subsequently, the method selects at step 215 a predictive model for performing classification based on a result of the evaluation cycle. Thus, the selection may be based on a comparison of the AUCs of the plurality of predictive models. Accordingly, the selected model may have a predictive power score that reaches a predictive power threshold, or the highest predictive power score among the predictive models.

[0052] This selection may allow to select, for each of the plurality of feature selection functions, models with a statistically significant AUC (i.e. , with a confidence interval between 0.5 and 1 ) in both training and test sets. A Bayesian information criterion (BIC) may be used to identify the optimal model. During selection, the models having only one variable can be excluded.

[0053] The radiomic features which may be used to differentiate benign (UMs) from malignant (LMSs) neoplastic lesions when applied to predictive models, for each of the plurality of feature selection functions and according to BIC value, are reported in table 1 below. As will be appreciated, these features can be used as input parameters of a generalized linear model.Table 1

[0054] Other standard metrics can be used for selecting the predictive models. These metrics can include accuracy, sensitivity, and specificity which are defined as follows:TP TKJTP + TNSENSITIVITY = — TP —+FN SPECIFICITY = — TN+ —FP ACCURACY = TN+FP+TP+FN

[0055] Whereby the TP variable indicates the number of images of LMS correctly classified as comprising an LMS, the TN variable indicates the number of images of UM correctly classified as UM, the FP variable indicates the number of images of an UM incorrectly classified as LMS, and the FN variable indicates the number of images of LMS incorrectly classified as UM.

[0056] The method described above allows to obtain performance with accuracy of 0.75-0.83, sensitivity of 0.81 -0.85 and specificity of 0.88-0.94.

[0057] An advantage of the technical solutions of the present embodiments is to provide a method for training a system for classifying a lesion as UM or LMS and a method for classifying a lesion as UM or LMS, which provide for optimal and reliable discrimination between malignant and benign lesions of the uterus in a pre-operative setting and exceed the discriminatory capacity of experienced radiologists on the same images.

[0058] Another advantage of the present technical solutions is that the methods herein described provide the ability to diagnose tumours in a manner that reduces the possibility of false negative results and avoids the risk of neoplastic dissemination during the operatory procedure and delays in diagnosis.

[0059] Another advantage of the present technical solutions is to provide a methodfor classifying a lesion as UM or LMS based on CT image(s), which provides a high level of sensitivity without requiring additional imaging investigations.

[0060] While aspects of the invention have been described in terms of various specific embodiments, it will be apparent to those of ordinary skill in the art that many modifications, changes, and omissions are possible without departing form the spirit and scope of the claims. In addition, unless specified otherwise herein, the order or sequence of any process or method steps may be varied or re-sequenced according to alternative embodiments.

[0061] Reference has been made in detail to the embodiments of the disclosure, one or more examples of which are illustrated in the drawings. Each example is provided by way of explanation of the disclosure, not limitation of the disclosure. In fact, it will be apparent to those skilled in the art that various modifications and variations can be made in the present disclosure without departing from the scope or spirit of the disclosure. Reference throughout the specification to "one embodiment" or "an embodiment" or “some embodiments” means that the particular feature, structure or characteristic described in connection with an embodiment is included in at least one embodiment of the subject matter disclosed. Thus, the appearance of the phrase "in one embodiment" or "in an embodiment" or "in some embodiments" in various places throughout the specification is not necessarily referring to the same embodiment(s). Further, the particular features, structures or characteristics may be combined in any suitable manner in one or more embodiments.

[0062] When elements of various embodiments are introduced, the articles “a”, “an”, “the”, and “said” are intended to mean that there are one or more of the elements. The terms “comprising”, “including”, and “having” are intended to be inclusive and mean that there may be additional elements other than the listed elements.REFERENCES1. D’Angelo, E.; Prat, J. Uterine Sarcomas: A Review. Gynecol. Oncol. 2010, 116, 131-139, doi:10.1016 / j.ygyno.2009.09.023.2. Ricci, S.; Stone, R.L.; Fader, A.N. Uterine Leiomyosarcoma: Epidemiology, Contemporary Treatment Strategies and the Impact of Uterine Morcellation.Gynecol. Oncol. 2017, 145, 208-216, doi: 10.1016 / j.ygyno.2017.02.019.3. Dondi, G.; Porcu, E.; De Palma, A.; et al. Uterine Preservation Treatments in Sarcomas: Oncological Problems and Reproductive Results: A Systematic Review. Cancers 2021, 13, 5808, doi: 10.3390 / cancersl 3225808.4. Rizzo, A.; Nannini, M.; Astolfi, A.; Indio, V.; De laco, P.; Perrone, A.M.; De Leo, A.; Incorvaia, L.; Di Scioscio, V.; Pantaleo, M.A. Impact of Chemotherapy in the Adjuvant Setting of Early Stage Uterine Leiomyosarcoma: A Systematic Review and Updated Meta-Analysis. Cancers 2020, 12, 1899, doi: 10.3390 / cancersl 2071899.5. Owen, C.; Armstrong, A.Y. Clinical Management of Leiomyoma. Obstet. Gynecol. Clin. North Am. 2015, 42, 67-85, doi:10.1016 / j.ogc.2014.09.009.6. Park, J.-Y.; Park, S.-K.; Kim, D.-Y.; Kim, J.-H.; Kim, Y.-M.; Kim, Y.-T.; Nam, J.-H. The Impact of Tumor Morcellation during Surgery on the Prognosis of Patients with Apparently Early Uterine Leiomyosarcoma. Gynecol. Oncol. 2011, 122, 255-259, doi: 10.1016 / j.ygyno.2011 .04.021 .7. Bretthauer, M.; Goderstad, J.M.; Loberg, M.; Emilsson, L.; Ye, W.; Adami, H.-O.; Kalager, M. Uterine Morcellation and Survival in Uterine Sarcomas. Ear. J. Cancer 2018, 101, 62-68, doi: 10.1016 / j.ejca.2O18.06.007.8. Valzacchi, G.M.R.; Rosas, P.; Uzal, M.; Gil, S.J.; Viglierchio, V.T. Incidence of Leiomyosarcoma at Surgery for Presumed Uterine Myomas in Different Age Groups. J. Minim. Invasive Gynecol. 2020, 27, 926-929, doi: 10.1016 / j.jmig.2O19.06.013.

Claims

CLAIMS1. A computer-implemented method (100) for classifying a lesion as a uterine leiomyoma, UM, or as a uterine leiomyosarcoma, LMS, the method (100) comprising: acquiring (101 ) at least one Computed Tomography, CT, image of the lesion; obtaining (103) a Region of Interest, ROI, of the lesion within the at least one CT image; calculating (105) one or more radiomic features of said ROI of the lesion; selecting (107), using a plurality of feature selection functions, a corresponding plurality of subsets of features, wherein each of said plurality of subsets of features comprises one or more selected radiomic features of said one or more radiomic features; applying (109) a corresponding plurality of predictive models which have been trained for classifying a lesion as UM or as LMS to the plurality of subsets of features; and classifying (113) the lesion as UM or as LMS using said plurality of the predictive models.

2. The method according to claim 1 , wherein acquiring (101 ) at least one CT image of the lesion comprises acquiring at least one contrast-enhanced computed tomography, CE-CT, image.

3. The method according to claim 1 or 2, further comprising processing (102) the at least one CT image to obtain at least one processed CT image having a preset isotropic voxel spacing and a pre-set voxel density; and wherein the obtaining (103) the ROI of the lesion is based on the least one processed CT image.

4. The method according to any one of the preceding claims, wherein said plurality of feature selection functions comprises a random forest optimizer function, a regression analysis function, and a recursive feature elimination function.

5. The method according to any one of the preceding claims, wherein said plurality of predictive models comprise machine learning-based models that have been trained and tested for the binary classification task based on results of a histological / histopathological examination of UM or LMS.

6. The method of any one of the preceding claims, wherein classifying (113) the lesion comprises: obtaining a probability that the lesion corresponds to UM or LMS based on the predictive model; and classifying the lesion as UM or LMS in response to a determination that said probability reaches a probability threshold value.

7. The method of any one of the preceding claims, further comprising performing a cross-correlation analysis (115) of the classification outputs obtained by using said plurality of predictive models; and determining (116) that the lesion is UM or LMS based on the cross-correlation analysis.

8. A computer-implemented method of training a system for classifying a lesion as a uterine leiomyoma, UM, or as a uterine leiomyosarcoma, LMS, the method (200) comprising: acquiring (201 ) a set of Computed Tomography, CT, images of lesions; obtaining (203) a Region of Interest, ROI, of the lesion within each of the set of CT images;acquiring (204) a set of labels, each being indicative that the corresponding ROI comprises a UM or a LMS; calculating (205) one or more radiomic features of said ROI of the lesion; selecting (207), using a plurality of feature selection functions, a corresponding plurality of subsets of features, wherein each of said plurality of subsets of features comprises one or more selected radiomic features of said one or more radiomic features; generating (209), based on the plurality of subsets of radiomic features, a plurality of predictive models that are trained for classifying a lesion as UM or as LMS based on the set of labels; and performing (211 ) a number of cross-validation cycles of the plurality of predictive models.

9. The method according to claim 8, wherein generating (209) a plurality of predictive models comprises combining the selected one or more radiomic features of each of the plurality of subset of radiomic features to fit a predictive model of said plurality of predictive models.

10. The method according to claim 9, wherein said predictive model is a generalized linear model.

11. The method according to any one of claims 8-10, wherein performing (211 ) a number of cross-validation cycles comprises, for each cycle and for each predictive model: adjusting one or more parameter values used to control the learning process of the predictive model; and estimating the predictive power score for classifying the lesion as UM or as LMS of the predictive model.

12. The method according to any one of claims 8-11 , wherein performing (211 ) a number of cross-validation cycles comprises, for each cycle and for each predictive model splitting the set of CT images into a training subset and test subset with balanced output.

13. The method according to any one of claims 8-12, further comprising: performing (213) an evaluation cycle of the predictive models; and selecting (215) a predictive model for performing classification based on a result of the evaluation cycle.

14. The method according to claim 13, wherein performing (213) an evaluation cycle comprises, for each of said predictive models: calculating an area, AUC, below a receiver operational characteristic curve, ROC, associated with the predictive model; and estimating the predictive power score of the predictive model based on said AUC.

15. The method according to claim 13, wherein selecting (215) a predictive model for performing classification comprises selecting (215) a predictive model having the predictive power score that reaches a predictive power threshold.

16. A system (1 ) comprising: a memory (10) for storage of at least one Computed Tomography, CT, image of a lesion; and a processing unit (20) communicatively coupled to the memory (10) and configured to execute the method of any one of the preceding claims.

17. A computer program product comprising one or more instructions that, when executed by a processing unit (20), cause a system (1 ) to carry out the method of any one of claims 1 -7 or 8-15.

18. A computer-readable storage medium comprising instructions which, when executed by a processing unit (20), cause a system (1 ) to carry out the method of any one of claims 1 -7 or 8-15.

Images