Method for predicting the dynamics of atopic dermatitis severity in at least one area of ​​a patient

Abstract

The invention relates to a computer-based method for predicting the dynamics of atopic dermatitis severity in at least one area of ​​a patient from time-series images of the patient's skin. The method comprises an inference phase including the following steps: (a) segmenting each image to delineate at least one atopic dermatitis lesion within the image, and (b) predicting the evolution over time of at least one severity element of the segmented lesion, using a Bayesian state-space model trained for predicting each severity element. Abstract figure: Figure 6

Description

Title of the invention: Method for predicting the dynamics of the severity of atopic dermatitis of at least one area of ​​a patient Technical field

[0001] The present invention relates to a method implemented by computer means for predicting the dynamics of severity of atopic dermatitis of at least one area of ​​a patient on the basis of time series images of the skin of said patient. Prior art

[0002] Atopic dermatitis (AD) or eczema is a prevalent chronic skin condition characterized by dry, itchy skin. The disease has periods of flare-ups and remissions, which are often unpredictable, making treatment difficult and burdensome for patients. Current treatments for mild and moderate AD involve the use of moisturizers on dry skin and anti-inflammatory creams or ointments (such as topical corticosteroids and calcineurin inhibitors) on inflamed skin. It is essential to individualize treatment approaches to each patient's specific conditions to achieve optimal efficacy, as treatment responses can vary between individuals and even within the same person.The development of personalized treatment strategies relies on the ability to predict the future state of the disease for each patient, as AD symptoms fluctuate in dynamic and highly varied ways.

[0003] The Harmonizing Outcome Measures for Eczema (HOME) initiative has suggested using the Eczema Area and Severity Index (EASI) as the primary measurement tool for assessing clinical signs of eczema in clinical trials. The SCORing AD (SCORAD) and its objective component (oSCORAD) have also been validated as outcome measures, while other scoring systems such as the Six Area Six Signs AD (SASSAD) are commonly used in clinical practice. These instruments provide a single score that reflects the overall severity of eczema by combining scores from various severity components, including intensity signs, subjective symptoms, and extent of the affected area. Each severity criterion captures a distinct aspect of eczema severity and may exhibit its own patterns of change over time.

[0004] Such scoring systems can be complex and time-consuming for practitioners to calculate. The present paper proposes a computer-implemented method for reliably predicting the dynamics of atopic dermatitis severity.

[0005] The state of the art to be considered can be found in: Suhendra, Rivansyah, et al. "A novel approach to multi-class atopy dermatitis disease severity scoring using multi-class SVM." 2019 IEEE International Conference on Cybemetics and Computational Intelligence (CybemeticsCom), IEEE, 2019; and Helske, Jouni, and Matti Vihola. "bssm: Bayesian Inference of Non-linear and Non-Gaussian State Space Models in R." arXiv preprint arXiv:2101.08492 (2021). Abstract

[0006] The present document proposes a method implemented by computer means for predicting the dynamics of severity of atopic dermatitis of at least one area of ​​a patient from time series images of the skin of said patient, said method comprising an inference phase comprising the following steps:

[0007] (a) segmenting each image to delineate at least one dermatitis lesion atopic in said image,

[0008] (b) predicting the evolution over time of at least one severity element of said segmented lesion, using a Bayesian state-space model trained for the prediction of each severity element.

[0009] The dynamics of atopic dermatitis (AD) severity refers to the patterns and changes in the severity of AD symptoms over time. As mentioned above, AD is known for its fluctuating nature, with periods of exacerbation (flare-ups) and periods of remission. Severity dynamics encompass variations in symptom intensity, duration, frequency, and extent of skin involvement throughout the disease course.

[0010] The dynamics of AD severity can vary considerably between individuals and even within the same person, as it is influenced by various factors such as genetic predisposition, environmental triggers, immune system responses, and individual patient characteristics. Understanding the dynamics of severity is essential for effective management and treatment of AD.

[0011] During flare-ups, the severity of AD symptoms tends to increase, with notable manifestations such as intense itching, redness, dryness, scaling, and skin lesions. These flare-ups may be triggered by factors such as allergens, irritants, stress, climate changes, or infections. The dynamics of severity may include rapid onset and progression of symptoms, resulting in significant discomfort and impairment of the patient's quality of life.

[0012] Conversely, during periods of remission, AD symptoms lessen or become less severe, resulting in improved skin condition and reduced symptoms. Remission phases may be spontaneous or occur with the help of treatments or lifestyle adjustments.

[0013] The dynamics of AD severity can vary in terms of the duration and frequency of flares and remissions. Some people experience frequent and prolonged flares with shorter periods of remission, while others may have longer periods of remission between flares. The severity of flares can also vary, ranging from mild to severe.

[0014] Monitoring and understanding the dynamics of AD severity is essential to tailor treatment strategies to each patient. This allows healthcare providers to adjust medications, lifestyle modifications, and preventive measures according to each patient's specific needs, with the aim of achieving optimal disease control, minimizing symptoms, and improving overall quality of life.

[0015] In machine learning and statistical modeling, the inference phase and the training phase are two distinct stages of the model life cycle.

[0016] During the training phase, the model may be trained on a set of labeled data to learn patterns, relationships, and parameters that enable it to make predictions or perform a specific task. The training data consists of known input examples and corresponding output labels or target values. The model iteratively adjusts its internal parameters based on the training data to minimize the deviation between the predicted outputs and the actual labels. This process is often referred to as optimization or learning. The training phase aims to find the best set of model parameters that generalize well to unseen data and enable accurate predictions.

[0017] In other words, the training phase is the stage in which a machine learning model learns patterns and relationships from labeled data in order to make accurate predictions or perform a specific task.

[0018] Once the model is trained, it can be used during the inference phase, also known as the deployment phase or the prediction phase. During this phase, the trained model is used to make predictions or perform the desired task on new, unseen data. The inference phase involves applying the trained model to the input data and obtaining predictions or output estimates based on the patterns and relationships it learned during training. Inference is typically the operational phase during which the model is used to provide real-time predictions or perform tasks in production or practical applications.

[0019] In other words, the inference phase is the step in which the trained model is used to make predictions or perform tasks on new, unseen data, in real time or in practical applications.

[0020] Time series images refer to a collection of images captured over a period of time, for example between 4 and 10 days, where each image represents a specific moment or frame in that sequence. These images are typically captured at regular intervals and are used to capture changes or variations that occur over time.

[0021] This image may be a 2D image. A 2D image is an image that exists in two dimensions and consists of rows and columns of pixels, where each pixel represents a specific color or intensity value. A 2D image has no depth or perception of three-dimensional space.

[0022] This image may depict a part of the patient's body, such as the head, trunk, arms, hands, legs, feet, genitals and / or skin in close-up.

[0023] Segmentation of a skin lesion in an image refers to the process of identifying and delineating the boundaries of the lesion relative to the surrounding healthy skin in the image. It may involve separating the region of interest, i.e., the skin lesion, from the background.

[0024] Segmentation is an important step in medical image analysis, particularly in dermatology, as it allows for the quantitative analysis and measurement of various lesion characteristics, such as size, shape, and texture. It also plays a crucial role in computer-assisted diagnosis and treatment planning.

[0025] The process of segmenting a skin lesion may generally include at least some of the following steps: - Preprocessing: The input image may undergo preprocessing steps such as noise removal, contrast enhancement, and normalization to improve image quality and uniformity. - Image enhancement: Techniques such as histogram equalization, filtering, and morphological operations can be applied to enhance features and highlight lesion boundaries. - Feature extraction: Image features relevant for skin lesion segmentation can be extracted. These can include information about color, texture, and local descriptors of the image. - Segmentation algorithm: Different image segmentation algorithms can be used, depending on the characteristics of the skin lesion and the specific requirements of the application. - Post-processing: After obtaining an initial segmentation, post-processing steps can be applied to refine and improve the segmentation result. These can include techniques such as morphological operations (erosion, dilation), region filtering, or contour smoothing.

[0026] The aforementioned segmentation algorithm may include thresholding, region-based segmentation (e.g., region growing or region splitting and merging), edge-based segmentation (e.g., edge detection, active contours (snakes) or level sets), machine learning-based segmentation (e.g., using convolutional neural networks (CNNs)).

[0027] As noted above, in the context of assessing the severity of atopic dermatitis (AD), various severity criteria or scoring systems may be used to assess the extent and intensity of the disease. These severity criteria are designed to provide a standardized and objective means of measuring the severity of atopic dermatitis symptoms and tracking changes over time.

[0028] A Bayesian state-space model (BSSM) is a statistical framework that combines state-space modeling with Bayesian inference. It can be used to analyze time-series data (here skin images over time) where the underlying process is assumed to have unobserved or latent states. BSSMs can be particularly useful when dealing with dynamic systems, such as biological processes, where states evolve over time and are not directly observable.

[0029] In a BSSM, the system can be represented by two components: the equation of state and the observation equation.

[0030] The equation of state can describe the evolution of latent states over time. It is generally represented as a set of equations that specify the transition dynamics of the states. The states can be continuous or discrete and represent various aspects of the system being modeled.

[0031] The observation equation can relate the observed data to the underlying states. It describes the relationship between the latent states and the observed variables. The observed variables can be noisy measurements of the latent states, and the observation equation models the measurement process.

[0032] The objective of a BSSM may be to estimate the unknown states from the observed data and to deduce the parameters of the state and observation equations. Bayesian inference can be used to obtain the posterior distribution of states and parameters, taking into account both prior knowledge and observed data.

[0033] Bayesian inference in BSSMs involves specifying prior distributions for model parameters and using Bayes' theorem to update the prior beliefs based on the observed data, yielding the posterior distribution. This posterior distribution can provide a probabilistic representation of unknown states and parameters as a function of the data, allowing uncertainty to be quantified.

[0034] The time period of this prediction can be between 4 and 10 days.

[0035] The method according to the present document may have the following advantages over the prior art: - Precise localization: Segmentation allows for precise delineation of the boundaries of atopic dermatitis lesions in the skin image. This precise localization allows for targeted analysis and prediction of severity dynamics, specifically within the identified lesions. It allows for the distinction of affected areas from unaffected areas, thus reducing the risk of erroneous predictions due to the inclusion of unrelated regions. - Objective and quantitative assessment: The use of computer-based procedures allows for an objective and quantitative assessment of the dynamics of atopic dermatitis severity. It eliminates the subjectivity or potential variability associated with manual assessments by providing standardized measures. - Targeted assessment: By segmenting atopic dermatitis lesions, the method focuses on specific areas of interest. It allows for a more targeted and detailed analysis of severity dynamics within lesions, providing specific information on how severity changes over time. This targeted assessment can be valuable in understanding disease progression, response to treatment, and localized variations in severity. - Non-invasive and convenient: The process relies on skin image analysis, which is a non-invasive and convenient approach for patients. It eliminates the need for physical examinations or invasive procedures, making the assessment process more comfortable for patients. - Increased accuracy and efficiency: Computerized algorithms improve the accuracy of severity assessment compared to manual processes. In addition, the use of machine learning techniques, such as than Bayesian state-space models, improves predictive capabilities, enabling more accurate and reliable predictions of severity dynamics. Segmentation also helps improve the performance of predicting severity dynamics. By isolating lesions, the method can account for the unique characteristics and dynamics specific to affected areas. The Bayesian state-space model, trained specifically on segmented lesion data, can better capture complex temporal patterns and provide more accurate predictions of severity evolution. - Real-time monitoring: The process allows for real-time monitoring of the dynamics of atopic dermatitis severity. By predicting the evolution of severity elements over time, clinicians can track disease progression and make informed decisions regarding treatment strategies and adjustments. - Personalized Therapeutic Approach: The method facilitates personalized therapeutic approaches by providing insight into the severity dynamics of specific areas in individual patients. This allows healthcare providers to tailor treatment plans based on predicted severity dynamics, optimizing the effectiveness of interventions and potentially reducing the burden of disease for patients. - Research and comparative analysis: Segmented lesions and predicted severity dynamics provide valuable data for research and comparative analysis. Researchers can analyze the dynamics of atopic dermatitis in specific lesion regions, evaluate the effectiveness of different treatments, and explore correlations between severity progression and various factors. This can contribute to the advancement of knowledge in this field and the development of new ideas and strategies.

[0036] Step (b) may consist of predicting the evolution over time of several severity elements of said segmented lesion, using a Bayesian state-space model for the prediction of each severity element.

[0037] Said method may comprise the following step:

[0038] (c) aggregating the prediction of each severity element made by each Bayesian state-space model to determine an aggregated severity score of said segmented lesion.

[0039] The aggregated severity score may be the SCORAD and the severity elements may include signs of extent and intensity of said segmented lesion. Said SCORAD may be based on each lesion of the patient.

[0040] Signs of intensity may include dryness, redness, swelling, crusting or oozing, scratching marks and / or thickening of said segmented lesion.

[0041] Severity elements may also include subjective symptoms.

[0042] Subjective symptoms may include itching and sleep disturbances.

[0043] Of course, other aggregated severity scores can be used, for example: - Eczema Area and Severity Index (EASI) - Patient-Oriented Eczema Measurement (POEM) - Investigator's Global Assessment (IGA) - Three-level severity (TIS) - Numerical Rating Scale (NRS)

[0044] The Bayesian state-space model for extent prediction can be a binomial Markov chain.

[0045] The Bayesian state-space model for predicting the extent of each intensity sign can be an ordered logistic random walk.

[0046] The Bayesian state-space model for predicting each subjective symptom can be a binomial random walk.

[0047] Bayesian state space models are defined below, with the following notations being used in this document: • ( t ) is the score of interest of the k-th patient at time f. k is deleted for models that do not take into account patient dependency. • M is the maximum value that can theoretically take, that is to say ye [0, M]. • y ~ N (fi, cr2) indicates that is normally distributed with mean and variance ^2. • v ~ N* (0, cr2) indicates that it follows a semi-normal distribution (normal distribution defined for positive values) with variance <72. • y ~ logN (fi, a2) indicates that is distributed in a log-normal way, that is, log(y) ~ N (fi, cr2). • y ~ logitN (g, a2) indicates that is distributed in a logit-normal way, that is, logit (y) ~ N (fi, a2 ). Range Models#: Binomial Markov Chain

[0048] The extent model assumes that the body surface can be subdivided into one hundred areas, each having a probability, p(y - 1 ) = p, of being classified as lesional, and that each area has fixed transition probabilities between the lesional and non-lesional states. The measure is specified as a distribution binomial to count the number of lesion areas to produce the extent score, A ~ B( 100, p).

[0049] Two-state Markov chain models for latent dynamics

[0050] The transition from the non-lesional state to the lesional state or from the lesional state to the non-lesional state for a given area of ​​the skin is modeled by a two-state Markov chain characterized by the following transition matrix [° 051 ] T= [ P ° 0 = P' P(il P(n I ^10 ^11 / \ P\Q

[0052] where • Pï0 is the probability of transition between the lesion state and the non-lesion state, i.e. the probability that a lesion area will be classified as non-lesion at the next stage, is the probability that a lesioned area will remain lesional and can be interpreted as a measure of the persistence of eczema, and • Pm is the transition probability from the non-lesional to the lesional state, i.e. the probability that a non-lesional area will be classified as lesional at the next stage, and can be interpreted as a measure of the susceptibility of eczema to develop symptoms. The probability, p(t) of a given area being lesional is calculated as follows P{t + 1) =pnp{t) + p^ï-pÇt) ).

[0053] The state of a skin area at time t is designated by xt = ( 1 0) if it is non-lesional and xt - (0 1 ) if it is lesional. The predictions of the state at time t + h are given by r Y ph r xt+h~ with 7T -n \ - (1-tt) 1-tf / '

[0054] where 2 = 1 -p - p is one of the eigenvalues ​​of T (the other is 1) and _ ui iu TT — n + n 1 L0 characterizes the steady-state (limit) distribution, X» =(1-^ jT) to which the Markov chain converges if the prediction horizon, h is sufficiently long. The value of 2 indicates the mobility of the Markov chain, i.e., the speed at which it converges to the steady-state distribution, with the value of 121 ^0 indicating a faster convergence.

[0055] We assume that the probabilities, ^oi and P^, are patient dependent and that either or both probabilities are time dependent, since the Markov chain converges to the steady state distribution, _ . which is Tl — n + n 2 01 7 10 likely to evolve over a sufficiently long period. We further assume that Pw = 1- (where / 'n is a measure of the persistence of eczema) does not present

[0056]

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068]

[0069]

[0070]

[0071] not have a strong time dependence and that Pqi (a measure of the susceptibility of eczema to develop symptoms) changes dynamically due to endogenous and exogenous factors such as skin barrier integrity and environmental stressors, respectively. We therefore parameterize the Markov chain by a time (t) and patient (k) dependent ^(t) and a patient dependent ^(t), from which we can infer = Pl0ltiï^- Prior data of the Markov chain binomial model The evolution of the extent, A®(f), for the ^-th patient at time 1 is modeled using a binomial measurement model and a latent Markov chain model described above (the model equations include ) since the limit upper limit of ( ) conditioned by n W is —1— where Pq] is between 0 and K 7 7 10 1+» 1): p® (t +1 ) = p + p 1 - / / *0' n (k) — X - n W ^11 ^10 a / 'io =pi(;kHS^ 1 w logit(7r{k\t+!) )-€ N€ ( logit(fr <k'(t) ), a2), avec les données préalables, N+(0, (0.251og(5) )2), p10 <feMogitN( / / 10, ^(f0)~logitN(-l,l), Z*1O~N(O,1), 1.52). The prior data are chosen to be weakly informative and result in reasonable prior predictive distributions. • The prior data on 17 results in an odd ratio increment for 77 of at most 5. That is, ^-( / )- 0,= 1 / 9) evolves to Æ( / + 1) = 0.36 (OMf + 1) = « 0:56), œ which can be considered an exceptionally important change.

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] • The prior data for the initial condition, , at the first data point are slightly biased towards 0, since large values ​​of A are very unlikely. • The prior data on and °3o translate into approximately uniform prior data on Intensity Sign Models: Ordered Logistic Random Walk It is assumed that the intensity sign measures are generated from an ordered logistic distribution, which is suitable for describing ordinal random variables with only a few categories. An ordered logistic distribution is a logistic distribution with a location parameter (latent score), which is integrated between M thresholds, = Q < < ... < ~ OrderedLogistic^^ c) l-logit4(y^(t) - c0) if i = Q = i) = logit'!(yg(t) ' Gi) ' logit^yj^t) -c{ ) for Q <i<M-t togit^yg^)- cMA) ifi = M. Q) = 0 is fixed without loss of generality since the ordered logistic distribution model is translation invariant (adding 2 to y^t etc does not change the distribution). The scale y^t depends on c, and is approximately cM-r co = ^wi. We therefore choose to express the prior data and latent dynamics with ^(t) = y£(t ) a normalized version of y^t. We assume that y® (t) follows a Gaussian random walk with variance <72. vRt+1) -batch and use the previous data: cro~N+(O,0,1252)' where 5 are the differences between consecutive thresholds, i.e. C; for ie [0, Ml],

[0086] The prior data are chosen to be weakly informative and result in reasonable prior predictive distributions. • The prior data on 5 results in values ​​less than the width of the standard logistic distribution, which allows us to consider the possibility of = i) ~ 1 (standard deviation of the logistic distribution standard is ). • The prior data on and ao translate to approximately uniform prior data for the initial latent score, ( tn 1 in the interval of • The prior data on ° assume that it is possible to move from a state where y = 0 is the most likely outcome, to a state where y = M is the most likely outcome, in two transitions. Subjective symptom models: Binomial random walk

[0087] The latent score, J® ( t ), for subjective symptoms (discrete variables with a large number of categories) is assumed to follow a random walk on the logit scale, and the measure is modeled by a binomial distribution whose success parameter is allowed to vary over time,

[0088] / *)( / )€ ~€ B(M, ^(t)\

[0089] logit(jX^(^+ 1)) ~€ N(logit( / X^(t) ), o2\

[0090] with the prior data,

[0091] <7~N+(0, (0.251og(5))2)'

[0092] ~iogi{^),

[0093] ^(o,?'

[0094]

[0095] The prior data are chosen to be weakly informative and result in reasonable prior predictive distributions. • The prior data on results in an odd ratio increment for p of at most 5. For example, if pd} — 0 1(£ZR( / )= §« 0 11 )' 'This prior data assumes that it is unusual, but not impossible, for the following value to be + 1) = 0.36 + 1) = 5OR(t) = 5 / 9 “0.56> • The prior data on and have a reasonable scope and result in marginal prior data for which are approximately uniform.

[0096] Step (a), or segmentation step, can be performed using a trained U-Net model.

[0097] U-Net is a convolutional neural network (CNN) architecture used in particular for image segmentation tasks.

[0098] The U-Net architecture is named after its U-shaped design, consisting of an encoding path (contraction path) and a decoding path (expansion path). The encoding path captures contextual information and reduces the spatial dimensions of the input image, while the decoding path recovers spatial information and generates a segmentation map.

[0099] The main characteristics of the U-Net architecture are as follows:

[0100] - Encoding Path: The encoding path consists of multiple layers of sub- sampling, typically implemented using convolutional and pooling operations. Each downsampling layer reduces the spatial dimensions while increasing the number of feature channels, allowing the network to capture increasingly abstract and contextual information.

[0101] - Expansion Path: The decoding path is responsible for recovering the spatial details and generating a high-resolution segmentation map. It consists of upsampling layers that increase spatial dimensions while reducing the number of feature channels. Upsampling layers can be accompanied by skip connections that concatenate feature maps from corresponding levels in the encoding path. These connections help preserve spatial details and combine them with contextual information.

[0102] - Jump Connections: U-Net's jump connections allow the flow of information between the encoding and decoding paths at multiple resolutions. By combining the high-resolution features of the encoding path with the features sampled in the decoding path, the network can efficiently localize and segment objects with fine detail.

[0103] - Symmetric architecture: U-Net's symmetric architecture ensures that the Spatial information lost during downsampling is effectively recovered during downsampling, resulting in accurate segmentation results.

[0104] This architecture is capable of capturing local and global contextual information while preserving the finest spatial details, making it particularly suitable for tasks requiring precise and accurate segmentation.

[0105] The Resnet model, for example ResNet-34, can be used as a backbone for the U-Net network.

[0106] In the context of the U-Net architecture, the backbone refers to the encoder portion of the network. The backbone typically consists of pre-trained convolutional neural network (CNN) models that have been shown to be effective in various computer vision tasks.

[0107] The choice of backbone for U-Net depends on the specific requirements of the task to be performed and the available pre-trained models. Backbones for U-Net include for example VGG, ResNet, MobileNet or EfficientNet.

[0108] Residual connections in the ResNet backbone help solve the vanishing gradient problem and facilitate learning deeper representations.

[0109] These backbones are typically used in transfer learning scenarios, where the pre-trained weights of the backbone are initialized with weights learned from large-scale image classification tasks such as ImageNet. The backbone is then fine-tuned to the specific image segmentation task using the U-Net architecture.

[0110] By using a pre-trained backbone as an encoder, U-Net can leverage the features learned during the large-scale classification task, allowing meaningful and high-level features to be extracted from the input image. This enables accurate and efficient segmentation of target objects or regions.

[0111] The present document also relates to a computer program comprising instructions for implementing the above-mentioned method, when this program is executed by a processor.

[0112] The present document also relates to a non-transitory computer-readable recording medium on which is recorded a program for implementing the above-mentioned method, when said program is executed by a processor.

[0113] This document also relates to a computing device comprising:

[0114] - an input interface for receiving time series images,

[0115] - a memory for storing at least the instructions of the computer program aforementioned,

[0116] - a processor accessing the memory to read the aforementioned instructions and then perform the above-mentioned process,

[0117] - an output interface for providing information based on the prediction of step (b). Brief description of the drawings

[0118] Other characteristics, details and advantages will appear on reading the detailed description below, and on analyzing the attached drawings, in which:

[0119] [Fig. 1] schematically represents an example of a computing device according to the present document,

[0120] [Fig.2] schematically shows an example of a method according to this document,

[0121] [Fig.3] illustrates (a) an original image of the skin of a first part of the body of the patient and (b) the segmentation masks of the surface of the lesions on this image,

[0122] [Fig.4] illustrates (a) an original image of the skin of a second part of the body of the patient and (b) the segmentation masks of the surface of the lesions on this image,

[0123] [Fig.5] illustrates the Bayesian state-space models according to this paper, where each model describes the dynamics of a latent severity (white ovals) and the measurement of the latent severity to obtain the recorded severity (grey ovals),

[0124] [Fig.6] illustrates the SCORAD prediction using nine models (rectangles colored), each of which corresponds to one of the nine SCORAD severity elements, which are aggregated to provide predictions for SCORAD. Description of the embodiments

[0125] [Fig.l] schematically illustrates an example of a computing device 1 according to the invention. Said computing device 1 comprises:

[0126] - an input interface 2,

[0127] - a memory 3 for storing at least the instructions of a program computer science,

[0128] - a processor 4 accessing the memory 3 to read said instructions and execute the process according to this document,

[0129] - an output interface 5.

[0130] The method according to the present document (as illustrated in [Fig.2]) is a method implemented by computer means for predicting the dynamics of severity of atopic dermatitis of at least one area of ​​a patient from time series images of the skin of said patient, said method comprising an inference phase comprising the following steps: - segmentation (SA) of each image to delineate at least one atopic dermatitis lesion in said image, using a trained segmentation model, more particularly a U-Net model, - the prediction (SB) of the evolution over time of the severity elements of said segmented lesion, using a Bayesian state-space model trained for the prediction of each severity element, - the aggregation (SC) of the prediction of each severity element made by each Bayesian state-space model to determine an aggregated severity score of said segmented lesion. Training and validation of the U-Net segmentation model# Datasets

[0131] Three annotated datasets were constructed to train and validate the performance of lesion surface segmentation. The first two datasets include only light-skinned patients (Fitzpatrick I-III), while the third dataset consists of images of dark-skinned patients (Fitzpatrick IV-VI).

[0132] The first dataset is a dataset collected from online dermatological atlases that consists of 604 images belonging to fair-skinned patients, one-third of whom are children, suffering from AD, with lesions present on different parts of the body. The dataset contains the following percentage of body areas: head (22%), trunk (11%), arms (23%), hands (9%), legs (16%), feet (8%), genitals (3%), full body (1%), and close-up skin (7%). The dataset contains a wide variety of clinical images taken with different angles, distances, light conditions, body parts, and disease severity. The images have a minimum size of 260 x 256 pixels, an average size of 667 x 563 pixels, and a maximum size of 1,772 x 1,304 pixels.

[0133] The second dataset was compiled for testing purposes. It was sourced from several publicly available dermatological atlases and contains a total of 367 images belonging exclusively to light-skinned patients. Each image was reviewed by a physician who approved its inclusion in the dataset. Duplicates or highly similar images were removed, and no further data sampling techniques were applied. The dataset contains images of children and adults with high variability in angles, distances, lighting conditions, body parts, and disease severity. The second dataset contains the following percentage of body areas: head (35%), trunk (20%), arms (18%), hands (7%), legs (13%), feet (2%), genitals (2%), and close-up skin (3%).Images have a minimum size of 313 x 210 pixels, a medium size of 574 x 537 pixels, and a maximum size of 2848 x 3252 pixels.

[0134] The third dataset is a dataset collected from online dermatological atlases that contain photos of children and adults with Fitzpatrick IV-VL skin types suffering from AD. It consists of 112 images with a minimum size of 200 x 204 pixels, an average size of 766 x 695 pixels, and a maximum size of 3,024 x 4,032 pixels. The dataset contains the following percentage of body areas: head (41%), trunk (10%), arms (17%), hands (8%), legs (13%), feet (3%), and close-up skin (8%). The purpose of including this dataset in the study was to gather preliminary evidence of the algorithm's effectiveness for dark skin. Ground truth marker

[0135] The corresponding ground truths of each dataset were prepared by nine experts, three for each dataset, who treat AD patients in their daily practice, to reduce variability by combining their results. The experts annotated the images without more context than the images. They had to draw a mask on the lesion. The ground truth markers were obtained for lesion segmentation by averaging the masks of the three annotators. The mean was preferred to the median because it is the statistical measure that gives the best results for generating ground truth markers from multi-annotator ordinal data. Data preprocessing

[0136] The images were resized to 512 x 512 and the pixel values ​​were scaled between 0 and 1. In addition, images in which the disease was too small were cropped, focusing on the disease. Ground truth markers were obtained by averaging the results, as explained in the previous section. Regarding the lesion surface masks, the average mask was calculated, resulting in a grayscale image between 0 and 255. An intensity threshold of 155 pixels was applied to obtain a binary mask that was used as ground truth. The images were finally normalized in the 0-1 range. Assessment measures

[0137] Pixel accuracy, area under the curve (AUC), intersection over union (loU), and Fl score were the preferred metrics to evaluate the segmentation model. Experimental setup

[0138] Two main experiments were conducted: one with images containing only light skin and the other by adding a small number of dark skin images to the training set. In the first experiment, the first dataset was used for training and the second and third datasets were used for testing. A six-fold cross-validation strategy was implemented to train the models. Models trained on the different levels were tested on both test sets, and the results were averaged across levels to reduce variance and bias.

[0139] The second experiment was performed to better understand the network's performance on dark skin by including a small fraction of dark-skinned patient images in the training set. In this experiment, the first dataset, the second dataset, and a subset of the third dataset were used for training and the rest for testing. The training and test subsets of the third dataset were obtained using a three-fold cross-validation strategy. This means that the training set consisted of 971 images of light-skinned patients (first and second datasets combined) and 75 images of dark-skinned patients (third dataset), which represents a tiny fraction of the total number of images (8%). The test set for dark skin consisted of the remaining 37 images from the third dataset. This split was performed three times (three levels), including different images in the training set and the test set, to obtain more reliable results. Results

[0140] The AUC, IoU and Fl for light skin were 0.93, 0.64 and 0.75 respectively, while the results of these measurements were 0.83, 0.32 and 0.42 respectively for dark skin.

[0141] When training on a small subset of dark skin images (experiment 2), it appears that the results are further improved (0.41 for IoU and 0.33 for Fl).

[0142] Figures 3 and 4 illustrate the segmented images for two different skin images. Training and Validation of Bayesian State-Space Models# Model Overview

[0143] A collection of nine machine learning models (Bayesian state-space models) is used to predict each of the nine severity features for SCORAD. The predictions from the nine models are aggregated to produce predictions for SCORAD assuming independence of the severity features. The latent dynamics and measurement distributions from the state-space models were fitted to each severity feature. The latent dynamics were modeled for extent as a Markov chain that describes how a small "area" of skin transitions from non-lesional to lesional and vice versa. A binomial distribution was used to count the number of lesional areas, i.e., extent. We assumed random walk latent dynamics for intensity signs and subjective symptoms in the absence of precise information on article-specific data generation mechanisms.The measurement of intensity signs from the latent scores was described by an ordered logistic measurement distribution and that of subjective symptoms by a binomial distribution. The parameters of the measurement distributions and the variance of the latent dynamics of the random walk were shared between patients, while the parameters of the latent dynamics of the Markov chain were made patient-dependent with . hierarchical prior data. The prior data were chosen to be weakly informative. Missing values ​​were treated as an absence of measurement in the state-space models.

[0144] Model inference was performed using the Hamiltonian Monte Carlo algorithm. Model validation

[0145] The predictive performance of the models was evaluated in a forward chaining approach, where the models were trained every four days. In other words, the models were first trained on data from the first day of a patient and tested using data from the next four days, then the models were trained on data from the first five days and tested on data from the next four days, and so on.

[0146] The probabilistic predictions of SCORAD and its nine severity elements can be evaluated using a logarithmic scoring rule, the logarithm of the predictive density (Ipd). A measure of accuracy can also be calculated for the SCORAD predictions.

Claims

Claims

1. A method implemented by computer means for predicting the severity dynamics of atopic dermatitis of at least one area of ​​a patient on the basis of time series images of the skin of said patient, said method comprising an inference phase characterized in that it comprises the following steps: a. segmenting each image to delineate at least one atopic dermatitis lesion in said image, b. predicting the evolution over time of at least one severity element of said segmented lesion, using a Bayesian state-space model trained for the prediction of each severity element.

2. Method according to the preceding claim, in which step (b) comprises predicting the evolution over time of several severity elements of said segmented lesion, using a Bayesian state-space model for the prediction of each severity element.

3. A method according to the preceding claim, wherein said method comprises the following step: (c) aggregating the prediction of each severity element made by each Bayesian state-space model to determine an aggregated severity score of said segmented lesion.

4. Method according to the preceding claim, in which the aggregated severity score is SCORAD and the severity elements comprise the signs of extent and intensity of said segmented lesion.

5. A method according to the preceding claim, wherein the signs of intensity include dryness, redness, swelling, crusting or oozing, scratch marks and / or thickening of said segmented lesion.

6. A method according to any preceding claim, wherein the time series images are captured over a time period of between 4 and 10 days.

7. A method according to any preceding claim, wherein the time period of said prediction is between 4 and 10 days.

8. A method according to any one of claims 4 to 7, wherein the Bayesian state-space model for extent prediction is a binomial Markov chain.

9. A method according to any one of claims 4 to 8, wherein the Bayesian state-space model for predicting each intensity sign extent is an ordered logistic random walk.

10. A method according to any one of claims 6 to 9, wherein the Bayesian state-space model for predicting each subjective symptom is a binomial random walk.

11. A method according to any preceding claim, wherein step (a) is performed using a trained U-Net model.

12. Computer program comprising instructions for implementing the method according to any one of claims 1 to 11, when this program is executed by a processor.

13. A non-transitory computer-readable recording medium on which is recorded a program for implementing the method according to any one of claims 1 to 11, when said program is executed by a processor.

14. A computer device (1) comprising: - an input interface (2) for receiving the time series image, - a memory (3) for storing at least the instructions of a computer program according to claim 12, - a processor (4) accessing the memory (3) to read said instructions and then execute the method according to any one of claims 1 to 11, - an output interface (5) for providing information based on the prediction of step (b).

Images