Detection of cutaneous melanoma by artificial intelligence

Abstract

The invention relates to a method for training (100) an artificial intelligence model (13) for detecting cutaneous melanoma, implemented by computer and comprising the following steps: - obtaining (101) images of skin having cutaneous melanoma and images of skin not having cutaneous melanoma, to form a training base containing these images; - training (104) the artificial intelligence model, from the training base, to provide numerical values corresponding to predictions relating to the absence or presence of cutaneous melanoma on respective skin images;- identification (112) of a threshold of absence or doubt, - identification (115) of a threshold of presence or doubt, so as to train the model so that, for a skin image provided as input to the trained model, the trained model provides as output a prediction of absence of a cutaneous melanoma on the image, a prediction of doubt on the presence or absence of cutaneous melanoma on the image, or a prediction of presence of a cutaneous melanoma on the image. Figure for abstract: figure 2;

Description

Title of the invention: Detection of cutaneous melanoma by artificial intelligence

[0001] The invention relates to the detection of a skin disease by an artificial intelligence model. It relates in particular to the detection of a cutaneous melanoma on a human subject from an image of the skin, by an artificial neural network.

[0002] Cutaneous melanomas were the sixth most common form of cancer in the European Union in 2020 and the twentieth most deadly form of cancer. The 5-year survival rate for the onset of cutaneous melanoma drops from 99% when detected early to 14% when detected at an advanced stage. It is therefore necessary to detect melanomas as early as possible.

[0003] However, in the case of a questionable skin lesion, it is sometimes long and difficult, even expensive, to obtain an appointment with a dermatologist to have the lesion examined. It is therefore necessary to simplify and speed up the process of obtaining an opinion on the presence or absence of cutaneous melanoma.

[0004] Artificial intelligence models are already known in the state of the art, in particular from document IN202241000096, which aim to provide a prediction on the presence or absence of a cutaneous melanoma on the skin of a human subject, from a photo of a skin lesion. However, while these models most of the time make correct predictions for extreme cases, i.e. clearly presenting either a malignant or a benign lesion, they still too often make errors for more questionable cases. However, it is crucial that the model does not provide a prediction of the absence of melanoma in a subject who is actually ill and vice versa.

[0005] The invention aims in particular to provide an artificial intelligence model making it possible to provide appropriate indications as to the presence or absence of a cutaneous melanoma.

[0006] To this end, the invention relates to a method for training an artificial intelligence model for detecting cutaneous melanoma, implemented by computer and comprising the following steps:

[0007] - obtaining images of skin presenting a cutaneous melanoma and images of skin not not presenting cutaneous melanoma, to form a learning base containing these images;

[0008] - training the artificial intelligence model, from the base learning, to provide numerical values ​​corresponding to predictions relating to the absence or presence of cutaneous melanoma on respective skin images;

[0009] - identification of a threshold of absence or doubt separating, on one side of this threshold, numerical values ​​provided by the trained model and corresponding to respective predictions of absence of cutaneous melanoma on respective skin images and, on the other side of this threshold, numerical values ​​provided by the trained model and corresponding to respective predictions of doubt on the presence or absence of cutaneous melanoma on respective skin images;

[0010] - identification of a threshold of presence or doubt separating, on one side of this threshold, the numerical values ​​provided by the trained model and corresponding to respective predictions of doubt on the presence or absence of cutaneous melanoma on respective skin images and, on the other side of this threshold, numerical values ​​provided by the trained model and corresponding to respective predictions of the presence of cutaneous melanoma on respective skin images,

[0011] so as to train the model so that, for a skin image provided as input to the trained model, the trained model provides as output a prediction of absence of a cutaneous melanoma in the image, a prediction of doubt about the presence or absence of cutaneous melanoma in the image, or a prediction of presence of a cutaneous melanoma in the image.

[0012] Thus, the model provides three possible predictions from a skin image including a skin lesion: it predicts the absence of cutaneous melanoma, in other words a low risk of melanoma, or it predicts a doubt about the presence or absence of melanoma, in other words a medium risk of melanoma, or it predicts the presence of melanoma, in other words a high risk of melanoma. The person learning the result can therefore guide the human subject towards appropriate care. If it is the human subject who learns the results themselves, they can inform the appropriate health professionals. Having three possible predictions instead of two, in particular having a possible prediction of doubt, makes it possible to avoid binary prediction errors with potentially serious consequences concerning “doubtful” cases.This in turn improves confidence in predictions of presence and in predictions of absence of melanoma.

[0013] Other optional features, taken alone or in combination, follow.

[0014] Advantageously, the method comprises, to identify the threshold of absence or doubt and the threshold of presence or doubt, the following steps:

[0015] - considering that a binary classification threshold separates, on one side of this threshold, numerical values ​​provided by the model and corresponding to respective predictions of the presence of cutaneous melanoma on respective skin images and, on the other side of this threshold, numerical values ​​provided by the model and corresponding to respective predictions of absence of cutaneous melanoma on respective skin images, for each binary classification threshold among a set of possible binary classification thresholds distinct from each other, determination of a precision value of the model, a recall value of the model, and an F-measure value of the model dependent on the precision and recall values ​​of the model,

[0016] - determination of the best value of F-measure of the model among the values ​​of F- determined measurements, this best value corresponding to the highest F-measurement value;

[0017] - choice of a tolerance percentage of the binary classification threshold;

[0018] - for each classification threshold among the set of classification thresholds binary, determining an equilibrium value corresponding to an absolute value of a difference between the best F-measure value and the result of a multiplication between the F-measure value corresponding to the binary classification threshold and the tolerance percentage, so as to form a set of equilibrium values, such that each binary classification threshold of the set of binary classification thresholds is associated with a respective equilibrium value of the set of equilibrium values;

[0019] - determination of a first local minimum of the set of equilibrium values;

[0020] - identification of the binary classification threshold associated with this first local minimum;

[0021] - identification of the threshold of absence or doubt as corresponding to this threshold associated with this first local minimum;

[0022] - determination of a second local minimum of the set of values of balance;

[0023] - identification of the binary classification threshold associated with this second minimum local ;

[0024] - identification of the presence or doubt threshold as corresponding to this associated threshold to this second local minimum.

[0025] Thus, considering that a binary classification threshold separates numerical values ​​provided by the model and corresponding to “positive” predictions (for example “presence of a cutaneous melanoma”) and “negative” predictions (“absence of a cutaneous melanoma”), we designate by “precision”, or “positive predictive value”, the proportion of correct positive predictions among the correct positive predictions and the correct negative predictions of the model, and we designate by “recall”, or “sensitivity”, the proportion of correct positive predictions among the correct positive predictions and the incorrect positive predictions.

[0026] Thus, to each possible binary classification threshold, among a set of thresholds, corresponds a precision value and a recall value of the model.

[0027] By "F-measure", or "F-score", or "Fp-score", we mean a measure of the performance of the model obtained on the basis of a precision value and a recall value of the model. It aims to ensure that the model makes correct positive predictions, which corresponds to a good recall, and in sufficiently large numbers, which corresponds to a good precision. It is obtained by the following formula

[0028] [Math.sl] F = (1+ 02)* (precision * recall) / (fi2* precision + recall)

[0029] where [3>0 is a parameter decided by a user or supplier of the method and aimed at determining the importance given to precision with respect to recall and vice versa. Thus, when [3=1, providing recall and precision with the same importance, the F-measure, which can be called "Fl-score", becomes

[0030] [Math.s2] F- 2* (precision * recall) / (precision + recall)

[0031] The F-measure belongs to the interval [0, 1].

[0032] Thus, each possible binary classification threshold corresponds to an F-measure. To identify the two thresholds of the invention, a supplier of the method chooses a percentage tolerance of the best F-measure, and the difference between each F-measure and the value corresponding to the percentage tolerance of the best F-measure is calculated. A curve of equilibrium values ​​with two minima is obtained. These two minima therefore correspond to the two F-measure values ​​closest to the best F-measure value to which a percentage tolerance has been added. These two values ​​provide by association the two thresholds of the invention. In other words, instead of having a single binary classification threshold, in particular a binary classification threshold corresponding to the highest F-measure, two thresholds are determined from this highest F-measure which correspond to F-measurements slightly higher and lower than this F-measurement according to a determined percentage tolerance.These two thresholds frame the numerical values ​​of the doubt predictions.

[0033] By carrying out these steps, we therefore ensure that we provide a doubt prediction for numerical values ​​that would be too close to the initial binary classification threshold. We therefore reinforce the security and confidence in the model. We also ensure that it remains useful by predicting a modest percentage, to continue to provide a sufficient number of absence and presence predictions.

[0034] Preferably, the method comprises, before the learning step, a step of increasing the learning base, by creating new images training images showing cutaneous melanoma from the training images, from the training base, showing cutaneous melanoma, preferably so as to have in the training base substantially 50% of training images showing cutaneous melanoma and substantially 50% of training images not showing cutaneous melanoma.

[0035] Thus, starting from images, other images are created to ensure a good representation of the “minority class”: images presenting cutaneous melanomas. This step therefore makes it possible to train the model on more images presenting cutaneous melanomas and therefore to improve the performance of the model.

[0036] Advantageously, the step of increasing the learning base comprises the following sub-steps:

[0037] - determination of a ratio value between the number of training images presenting a cutaneous melanoma and the number of training images not presenting a cutaneous image, in the training base, so as to determine a number of modified copies to be created from each training image in the training base presenting a cutaneous melanoma,

[0038] - for each training image showing a cutaneous melanoma, creation of modified copies, according to the determined number of copies, to create the new training images comprising cutaneous melanomas, by carrying out, to create each copy, at least one of the following transformations of the training image:

[0039] ** cropping the image so as to keep only one area in the copy of interest;

[0040] ** vertical and / or horizontal flipping of the image;

[0041] ** transposition of the image;

[0042] ** increasing image contrast using an equalization technique contrast-limited adaptive histogram;

[0043] ** random modification of a brightness and / or a saturation and / or of the image;

[0044] ** integration into the image of a Gaussian and / or directional and / or median blur;

[0045] ** integration, by random mask of the image, of Gaussian noise;

[0046] ** simulation, by several transformations of the image, of a photo of the image performed by microscope.

[0047] Thus, these transformations make it possible to create, in a simple manner, new learning images that are sufficiently distinct from the respective images on which they are based to allow relevant training of the model.

[0048] These transformations are isometric. Thus, they do not distort the initial training image that is copied.

[0049] Preferably, creating a copy of a training image also includes, as a transformation, detecting one or more hairs or bristles. in the image and its or their removal by an inpainting technique based on a method called fast marching.

[0050] The invention also relates to an artificial intelligence model, trained by the training method described above and configured to provide, from a skin image, a prediction of the presence of a cutaneous melanoma, the absence of cutaneous melanoma, or a doubt about the presence or absence of cutaneous melanoma, on the skin image.

[0051] Advantageously, the model comprises an artificial neural network, preferably a convolutional neural network.

[0052] Thus, these networks are the most suitable networks for image classification.

[0053] Preferably, the artificial neural network has an “EfficienNet” type architecture.

[0054] Indeed, EfficientNet implements a compound scaling principle, which adjusts the depth, width and resolution of the neural network in a balanced manner, resulting in better performance, in particular better accuracy, than conventional convolutional networks. The scaling principle also provides great flexibility to adapt the model to different resource constraints, for example its use on a mobile terminal, while maintaining high performance. The EfficientNetV2 version, the second version of the EfficientNet-like architecture, is more efficient than the previous version.

[0055] The invention also relates to a method for detecting a cutaneous melanoma, comprising the following steps:

[0056] - obtaining a skin image;

[0057] - prediction, by the artificial intelligence model previously described, of a presence of cutaneous melanoma, absence of cutaneous melanoma, or doubt about the presence or absence of cutaneous melanoma, on the skin image obtained.

[0058] The invention also relates to a melanoma detection device, comprising:

[0059] - an apparatus for obtaining an image of the skin of a human subject;

[0060] - a prediction system configured to provide a prediction, by the model artificial intelligence described above, of the presence of a cutaneous melanoma, of an absence of cutaneous melanoma, or of a doubt about the presence or absence of cutaneous melanoma, on a skin image obtained by the obtaining device.

[0061] Advantageously, the device comprises a web application, accessible on the web, configured to provide the prediction.

[0062] Thus, it is the web application that implements the detection method. It is accessible from any type of terminal, in particular from any type of terminal equipped with means of obtaining an image, such as a smartphone, a PC with a webcam, a tablet.

[0063] The invention also relates to a mobile terminal, comprising the detection device described above. This is for example a tablet or a smartphone.

[0064] The invention also relates to a data processing system comprising means for implementing the steps of the training method described above or the steps of the detection method described above.

[0065] The invention also relates to a computer program comprising instructions which, when the program is executed by a computer, cause the latter to implement the steps of the training method described above or the steps of the detection method described above.

[0066] The invention also relates to a computer-readable recording medium comprising instructions which, when executed by a computer, cause the latter to implement the steps of the previously described formation method or the steps of the previously described detection method. Brief description of the figures

[0067] The invention will be better understood on reading the following description, given solely by way of example and with reference to the appended drawings in which:

[0068] [Fig-1] is a diagram of an environment for implementing the invention;

[0069] [Fig.2] is a flowchart of a method for training a detection model of melanoma according to the invention;

[0070] [Fig.3] is a flowchart of a method for detecting a melanoma according to the invention. Detailed description

[0071] [Fig.l] shows a human subject 1 presenting an arm 3 comprising a skin lesion 5. A doctor 7 manipulates a mobile terminal 9.

[0072] The mobile terminal 9 is connected, by wireless connection means, to a training system 11, located on a remote server, of an artificial intelligence model 13 for detecting cutaneous melanoma, so that the system 11 transmits the trained model to the mobile terminal 9.

[0073] The model training system 11 is a data processing system comprising means, in particular a processor 15, for implementing the steps of the training method 100 described below to train the model 13. To do this, the system comprises, in a computer-readable recording medium 17, a computer program 19 comprising instructions which, when executed by a computer, here by the processor 15, cause the latter to implement implements the steps of the training method 100. The support 17 in other words itself contains instructions which, when executed by a computer, lead the latter to implement the steps of the training method 100. The system 11 also comprises a training database 21 comprising images of skin with cutaneous melanoma and images of skin without cutaneous melanoma.

[0074] The artificial intelligence model 13, trained by the system 11 in accordance with the method 100 described below, is a convolutional neural network corresponding to the EfficientNetV2 architecture, belonging to the EfficientNet architecture type. Alternatively, it could be another type of artificial neural network, or even another type of model. However, convolutional networks and particularly EfficientNet type networks prove advantageous in the context of the invention. This model 13 is configured to provide, from a skin image, a prediction of the presence of a cutaneous melanoma, the absence of cutaneous melanoma, or a doubt about the presence or absence of cutaneous melanoma, on the skin image. It is implemented in the context of the melanoma detection method 200 described below.

[0075] The mobile terminal 9 is a conventional smartphone. Some of the components it incorporates form a melanoma detection device 22.

[0076] This melanoma detection device 22 comprises an apparatus 23 for obtaining a skin image to take a photograph of the lesion of the human subject 1. This apparatus 23 is a conventional smartphone camera.

[0077] The detection device 22 also contains a prediction system 25 configured to provide a prediction, by the artificial intelligence model 13, of a presence of a cutaneous melanoma, an absence of cutaneous melanoma, or a doubt about the presence or absence of cutaneous melanoma, on the skin image obtained by the skin image obtaining apparatus 15. The prediction system 25 is a data processing system comprising means, in particular a processor 27, for implementing the steps of the detection method 200 described below. To do this, the system 25 comprises, in a computer-readable medium 29, a computer program 31 comprising instructions which, when the program is executed by a computer, here by the processor 27, lead the latter to implement the steps of the detection method 200.The computer program 31 is previously downloaded in the form of a software application 33 downloadable from a server. In other words, in the mobile terminal 9, this system 25 applies the model 13 through the software application 33, which implements the steps of the detection method 200, to detect a cutaneous melanoma. Alternatively, the application 33 is a web application, accessible via the device 22 from a remote server. The program 31 is then executed on the cloud through this application 33.

[0078] We will now describe the method 100 for training the model 13. It is implemented by the system 11.

[0079] Step 101 is the obtaining of skin images presenting a cutaneous melanoma and skin images not presenting a cutaneous melanoma, to form the training base 17 containing these images. These images are obtained from conventional medical data sets for those skilled in the art and number 60,000 images in total. The images are of the “dermoscopic” or “dermoscopic” type. These images could be of a different number. The manner in which the data set is separated into training data and validation data is conventional and will not be described.

[0080] Steps 102 and 103 are the increase of the learning base, by creating new learning images presenting a cutaneous melanoma from the learning images, of the learning base, presenting a cutaneous melanoma, so as to have in the learning base substantially 50% of learning images presenting a cutaneous melanoma and substantially 50% of learning images not presenting a cutaneous melanoma.

[0081] To do this, step 102 is the determination of a ratio value between the number of training images presenting a cutaneous melanoma and the number of training images not presenting a cutaneous melanoma, in the training base, so as to determine a number of modified copies to be created from each training image of the training base presenting a cutaneous melanoma. Here, this ratio is 0.1, that is to say that 10% of the images of the training base present a melanoma. It is therefore necessary to make nine modified copies of each of these images, to obtain a base presenting substantially 50% of images with a melanoma and 50% of images without melanoma.

[0082] Step 103 is thus, for each training image presenting a cutaneous melanoma, the creation of modified copies, according to the number of copies determined, here nine, to create the new training images comprising cutaneous melanomas, by carrying out, to create each copy, at least one of the following transformations of the training image:

[0083] - cropping the image so as to keep only an area of ​​interest in the copy;

[0084] -vertical and / or horizontal flipping of the image;

[0085] -transposition of the image;

[0086] - increasing image contrast using a contrast-limited adaptive histogram equalization technique (or “Contrast Limited Adaptive Histogram Equalization”);

[0087] -random modification of a brightness and / or a saturation and / or of the image;

[0088] -integration into the image of a Gaussian and / or directional and / or median blur;

[0089] - integration, by random mask of the image, of Gaussian noise;

[0090] -simulation, by several transformations of the image, of a photo of the image produced by microscope;

[0091] - detection of one or more hairs or bristles in the image and its or their deletion by an inpainting technique (which can be called "incrustation method" in French) based on a method called "fast marching" (which can be called "fast step method" in French).

[0092] Step 104 is the training of the artificial intelligence model 13, from the learning base, to provide numerical values ​​corresponding to predictions relating to the absence or presence of cutaneous melanoma on respective skin images. This training is supervised and carried out in a conventional manner. The manner in which the training is carried out is conventional and is not the subject of this application. This is also the case for the validation and calibration steps.

[0093] Once the training is finished, steps 105 to 114 allow the identification of an absence or doubt threshold separating, on one side of this threshold, numerical values ​​provided by the trained model 13 and corresponding to respective predictions of absence of cutaneous melanoma on respective skin images and, on the other side of this threshold, numerical values ​​provided by the trained model 13 and corresponding to respective predictions of doubt on the presence or absence of cutaneous melanoma on respective skin images, and these steps allow the identification of a presence or doubt threshold separating, on one side of this threshold, numerical values ​​provided by the trained model 13 and corresponding to respective predictions of doubt on the presence or absence of cutaneous melanoma on respective skin images and, on the other side of this threshold,numerical values ​​provided by the trained model 13 and corresponding to respective predictions of the presence of cutaneous melanoma on respective skin images, so as to train the model 13 such that, for a skin image provided as input to the trained model 13, the trained model provides as output a prediction of the absence of a cutaneous melanoma on the image, a prediction of doubt on the presence or absence of cutaneous melanoma on the image, or a prediction of the presence of a cutaneous melanoma on the image.

[0094] Thus, step 105 is, for each binary classification threshold among a set of possible binary classification thresholds distinct from each other, the determination of a precision value of the model, a recall value of the model, and an F-measure value of the model dependent on the precision and recall values ​​of the model, considering that a binary classification threshold separates, on one side of this threshold, numerical values ​​provided by the model and corresponding to respective predictions of the presence of cutaneous melanoma on respective images of skin and, on the other side of this threshold, numerical values ​​provided by the model and corresponding to respective predictions of absence of cutaneous melanoma on respective skin images.

[0095] The F measure, here called Fl_measure because it balances the importance of recall and precision, is determined based on the following formula:

[0096] [Math.s3] F\_measure = 2* (precision * recall) / (precision + recall)

[0097] where "precision" and "recall" are the precisions and recalls corresponding to the model for each classification threshold considered. We therefore have as many FL-measures as there are possible classification thresholds. We could also use the FB-Measure, providing more importance to the recall or precision depending on the parameter B according to the desired performance.

[0098] Step 106 is the determination of the best F-measure value of the model among the determined F-measure values, this best value corresponding to the highest F-measure value. Here it is 0.86 and corresponds to a binary classification threshold of 0.49.

[0099] Step 107 is the choice of a tolerance percentage. This percentage is chosen here to be 5%, but it could be different.

[0100] Step 108 is, for each binary classification threshold among the set of binary classification thresholds, the determination of an equilibrium value corresponding to an absolute value of a difference between the best F-measure value and the result of a multiplication between the F-measure value corresponding to the binary classification threshold and the tolerance percentage, so as to form a set of equilibrium values, such that each binary classification threshold of the set of binary classification thresholds is associated with a respective equilibrium value of the set of equilibrium values. Each equilibrium value is therefore determined by the following formula, for each F-measure corresponds to a classification threshold:

[0101] [Math.s4] Best F _measure - percentage * F measure) •>

[0102] where Abs is the absolute value, percentage the chosen percentage (here 5%).

[0103] We then have a curve of equilibrium values.

[0104] Step 109 is the determination of a first local minimum of the set of equilibrium values, that is to say a first minimum of the curve.

[0105] Step 110 is the identification of the binary classification threshold associated with this first local minimum. Here it is 0.37.

[0106] Step 111 is the identification of the absence or doubt threshold as corresponding to this threshold associated with this first local minimum.

[0107] Step 112 is the determination of a second local minimum of the set of equilibrium values.

[0108] Step 113 is the identification of the binary classification threshold associated with this second local minimum. Here it is 0.59.

[0109] Step 114 is the identification of the presence or doubt threshold as corresponding to this threshold associated with this second local minimum.

[0110] Thus, instead of a single binary classification threshold of 0.49, separating on one side the values ​​corresponding to a prediction of absence of melanoma and on the other side the values ​​corresponding to predictions of presence of melanoma, the invention makes it possible to have two thresholds, one separating on one side the values ​​lower than 0.37, corresponding to predictions of absence of melanoma, and on the other side values ​​greater than or equal to 0.37, the other threshold separating on one side the values ​​greater than or equal to 0.59, corresponding to predictions of presence of melanoma, and on the other side values ​​lower than 0.59, so that the values ​​located between 0.37 and 0.59 correspond to predictions of doubt on the absence or presence of melanoma. In other words, for values ​​too close to the binary classification threshold, the model provides a doubt prediction using the two thresholds of the invention.

[0111] The trained neural network, associated with these two thresholds, forms the artificial intelligence model 13 obtained by the method 100. This model 13 is therefore now configured to provide, from a skin image, a prediction of the presence of a cutaneous melanoma, the absence of cutaneous melanoma, or a doubt about the presence or absence of cutaneous melanoma, on the skin image.

[0112] Naturally, all the values ​​mentioned could be different.

[0113] We will now describe the method 200 for detecting a cutaneous melanoma. It is implemented by the mobile terminal 9, using the model 13 included in the terminal. Indeed, in a step not illustrated, the training system 11 transmits the trained model to the terminal. As a variant, this transmission is carried out regularly, during updates of the model 13, generally during an update of the application 33. In the variant according to which the application 33 is a web application, the model does not have to be transmitted to the terminal, it is executed on the cloud.

[0114] Step 201 is obtaining a skin image of the human subject. The doctor takes a picture of a skin lesion of the human subject, for example a mole that has appeared or which seems to have evolved recently. This is achieved by the camera of the mobile terminal 9.

[0115] Step 202 is the prediction, by the artificial intelligence model 13, of a presence of a cutaneous melanoma, an absence of cutaneous melanoma, or a doubt about the presence or absence of cutaneous melanoma, on the skin image obtained. Thus, thanks to the trained model and the identification of the two thresholds, the model 13 provides the doctor with a low, medium or high risk of cutaneous melanoma, independently of the doctor's own examination. This result can be used to guide the human subject towards a specialist, for example if the doctor is a general practitioner, or directly towards an appropriate treatment offer if necessary. Conversely, the result can reassure the human subject.

[0116] The invention is not limited to the embodiments presented and other embodiments will become clear to those skilled in the art.

[0117] In particular, a single system can combine the functions of model training and melanoma detection by the model. For example, if the model is not transmitted by the remote server on which it is trained to the mobile terminal 9, the software application 33 of the terminal can send a request to the model located on a remote server, containing the photo, while the model returns the prediction to the application.

[0118] Instead of a mobile terminal, any device having a means of capturing a skin image can be used. For example, a webcam can be used to capture the skin image, then sent to the model.

[0119] Similarly, the application 33 can be any type of software, web application or mobile application. List of references

[0120] 1: human subject 3: arm of the human subject 5: skin lesion

[0121] 7: doctor

[0122] 9: mobile terminal

[0123] 11: data processing system

[0124] 13: artificial intelligence model

[0125] 15: processor

[0126] 17: recording medium

[0127] 19: computer program

[0128] 21: learning base

[0129] 22: melanoma detection device

[0130] 23: camera

[0131]

[0132]

[0133]

[0134]

[0135]

[0136]

[0137] 25: data processing system 27: processor 29: recording medium 31: computer program 33: mobile or web application 100: method for training a melanoma detection model 200: method for detecting a melanoma

Claims

Claims

1. A method of training (100) an artificial intelligence model (13) for detecting cutaneous melanoma, implemented by computer and characterized in that it comprises the following steps: - obtaining (101) images of skin presenting a cutaneous melanoma and images of skin not presenting a cutaneous melanoma, to form a learning base containing these images; - training (104) the artificial intelligence model, from the learning base, to provide numerical values ​​corresponding to predictions relating to the absence or presence of cutaneous melanoma on respective skin images;- identification (112) of a threshold of absence or doubt separating, on one side of this threshold, numerical values ​​provided by the trained model and corresponding to respective predictions of absence of cutaneous melanoma on respective skin images and, on the other side of this threshold, numerical values ​​provided by the trained model and corresponding to respective predictions of doubt on the presence or absence of cutaneous melanoma on respective skin images;- identification (115) of a presence or doubt threshold separating, on one side of this threshold, the numerical values ​​provided by the trained model and corresponding to respective predictions of doubt on the presence or absence of cutaneous melanoma on respective skin images and, on the other side of this threshold, numerical values ​​provided by the trained model and corresponding to respective predictions of presence of cutaneous melanoma on respective skin images, so as to train the model so that, for a skin image provided as input to the trained model, the trained model provides as output a prediction of absence of a cutaneous melanoma on the image, a prediction of doubt on the presence or absence of cutaneous melanoma on the image, or a prediction of presence of a cutaneous melanoma on the image.;

2. Method (100) according to the preceding claim, comprising, to identify the threshold of absence or doubt and the threshold of presence or doubt, the following steps: - considering that a binary classification threshold separates, on one side of this threshold, numerical values ​​provided by the model and corresponding to respective predictions of the presence of cutaneous melanoma on respective skin images and, on the other side of this threshold, numerical values ​​provided by the model and corresponding to respective predictions of the absence of cutaneous melanoma on respective skin images, for each binary classification threshold from a set of possible binary classification thresholds distinct from each other, determining (105) a precision value of the model, a recall value of the model, and an F-measure value of the model dependent on the precision and recall values ​​of the model, - determining (106) the best F-measure value of the model from among the determined F-measure values, this best value corresponding to the highest F-measure value; - choice (107) of a tolerance percentage; - for each binary classification threshold among the set of binary classification thresholds, determining (108) an equilibrium value corresponding to an absolute value of a difference between the best F-measure value and the result of a multiplication between the F-measure value corresponding to the binary classification threshold and the tolerance percentage, so as to form a set of equilibrium values, such that each binary classification threshold of the set of binary classification thresholds is associated with a respective equilibrium value of the set of equilibrium values; - determination (109) of a first local minimum of the set of equilibrium values; - identification (110) of the binary classification threshold associated with this first local minimum; - identification (111) of the threshold of absence or doubt as corresponding to this threshold associated with this first local minimum; - determination (112) of a second local minimum of the set of equilibrium values; - identification (113) of the binary classification threshold associated with this second local minimum; - identification (114) of the presence or doubt threshold as corresponding to this threshold associated with this second local minimum.

3. Method (100) according to any one of the preceding claims, comprising, before the learning step, a step of increasing (102, 103) the learning base, by creating new learning images presenting a cutaneous melanoma from the learning images, of the learning base, presenting a cutaneous melanoma, preferably so as to have in the learning base substantially 50% of learning images presenting a cutaneous melanoma and substantially 50% of learning images not presenting a cutaneous melanoma.

4. Method (100) according to the preceding claim, in which the step of increasing the learning base comprises the following sub-steps: - determining (102) a ratio value between the number of learning images presenting a cutaneous melanoma and the number of learning images not presenting a cutaneous image, in the learning base, so as to determine a number of modified copies to be created from each learning image of the learning base presenting a cutaneous melanoma, - for each learning image presenting a cutaneous melanoma, creating (103) modified copies, according to the number of copies determined, to create the new learning images comprising cutaneous melanomas, by carrying out, to create each copy, at least one of the following transformations of the learning image: ** cropping the image so as to keep only an area of ​​interest in the copy;** vertical and / or horizontal flipping of the image; ** transposition of the image; ** increase of an image contrast from an adaptive histogram equalization technique limited by contrast; ** random modification of a brightness and / or a saturation and / or of the image; ** integration into the image of a Gaussian and / or directional and / or median blur; ** integration, by random mask of the image, of a Gaussian noise; ** simulation, by several transformations of the image, of a photo of the image taken by microscope.;

5. Method (100) according to the preceding claim, in which the creation (103) of a copy of a training image also comprises, as transformation, the detection of one or more hairs or bristles in the image and its or their removal by an inpainting technique based on a method called defast marching.

6. Artificial intelligence model (13), characterized in that it is formed by the training method (100) according to any one of the preceding claims and is configured to provide, from a skin image, a prediction of the presence of a cutaneous melanoma, the absence of cutaneous melanoma, or a doubt about the presence or absence of cutaneous melanoma, on the skin image.

7. Model (13) according to the preceding claim, comprising an artificial neural network, preferably a convolutional neural network.

8. Model (13) according to the preceding claim, in which the artificial neural network has an architecture of the “EfficientNet” type

9. ». Method for detecting (200) in vitro a cutaneous melanoma, characterized in that it comprises the following step: - prediction (202), by the artificial intelligence model (13) of any one of claims 6 to 8, of a presence of a cutaneous melanoma, of an absence of cutaneous melanoma, or of a doubt on the presence or absence of cutaneous melanoma, on a skin image previously obtained from an individual.

10. Melanoma detection device (22), characterized in that it comprises: - an apparatus for obtaining (23) a skin image of a human subject; - a prediction system (25) configured to provide a prediction, by the artificial intelligence model (13) of any one of claims 8 to 10, of a presence of a cutaneous melanoma, of an absence of cutaneous melanoma, or of a doubt about the presence or absence of cutaneous melanoma, on a skin image obtained by the obtaining apparatus.

11. Device (22) according to the preceding claim, comprising a web application (33), accessible on the web, configured to provide the prediction.

12.

13.

14.

15. Mobile terminal (9), comprising the detection device (22) of claim 10 or 11. A computer program (19, 31) comprising instructions which, when the program is executed by a computer, cause the computer to implement the steps of the training method (100) according to any one of claims 1 to 5 or the steps of the detection method (200) according to claim 9. Data processing system (11, 25) comprising means for implementing the steps of the training method (100) according to any one of claims 1 to 5 or the steps of the detection method (200) according to claim 9. A computer-readable recording medium (17, 29) comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the forming method (100) according to any one of claims 1 to 5 or the steps of the detecting method (200) according to claim 9.

Images