Assistance in the assessment of acute respiratory distress syndrome

Abstract

The present invention relates to assisting the assessment of acute respiratory distress syndrome (ARDS) in a person. Subject of the present disclosure are a computer-implemented method, a computer system, and a computer program product to obtain information to support the assessment of ARDS.

Description

[0001] Assistance in the Assessment of Acute Respiratory Distress Syndrome

[0002] The present invention relates to assisting the assessment of acute respiratory distress syndrome (ARDS) in a person. Subject of the present disclosure are a computer-implemented method, a computer system, and a computer program product to obtain information to support the assessment of ARDS.

[0003] ARDS is a life-threatening disease in which the lungs cannot work properly. ARDS is caused by damage in the capillary wall that can be attributed to a disease or physical injury. This damage makes the capillary wall leaky, which leads to an accumulation of fluid and ultimately to the collapse of the alveoli. As a result, the lungs are no longer capable of exchanging oxygen and carbon dioxide. ARDS does not usually occur as an independent disease but is the result of another disease or a severe accident or injury.

[0004] Although every tenth intensive care patient suffers from acute respiratory distress syndrome, this life-threatening complication is frequently not detected (JAMA, 315: 788-800, 2016;). Impending death can frequently be prevented through several simple interventions; they include early mechanical ventilation with positive end-expiratory pressure (PEEP) and a reduced tidal volume. A prone position of the patient is also recommended. In clinical studies, these measures distinctly reduced mortality in some cases. However, these recommendations can only be implemented if ARDS is detected early.

[0005] Artificial Intelligence-aided diagnosis for ARDS has been described by Kai-Chih Pai et al., 2022, Digital Health, Vol. 8, 1- 15. The authors used a convolutional neural network (CNN) model to determine the probability of ARDS with chest X-ray data and clinical feature data and obtained accuracy of the diagnosis of 80% (with x-ray data alone) and 85% (with x-ray data and clinical feature data). Nesrine Zaglam et al., 2014, Computers in Biology and Medicine, 52, 41-48 describe a computer-aided diagnosis system for ARDS from chest radiographs which uses a semi-automatic patch extraction including a manual bone annotation and an automatic sampling.

[0006] Proceeding from the prior art, it is an object of the present disclosure to provide a solution for providing information to support the assessment of the presence of ARDS in a person.

[0007] This object is achieved by the subject of the independent claims. Preferred embodiments can be found in the dependent claims and in the present description and the drawings. The present disclosure provides in a first aspect a computer-implemented method to obtain information to support the assessment of the presence of Acute Respiratory Distress Syndrome (ARDS) in a person, comprising the following steps:

[0008] - a) receiving and / or retrieving one or more radiologic image(s) of the chest of a person;

[0009] - b) processing the one or more radiologic image(s) of the chest of the person by: bl) segmenting the one or more radiologic image(s) to identify different zones within the chest region of said one or more radiologic image(s), b2) applying a bone suppression algorithm to the one or more radiologic image(s) of the chest of the person to suppress bones in the one or more radiologic image(s), b3) applying a feature extraction algorithm to the data obtained in steps bl) and b2);

[0010] - c) calculating a probability of the presence of ARDS with the data obtained in step b3) by applying a classification and / or a regression algorithm;

[0011] - d) output the information obtained in step c).

[0012] The present disclosure further provides a computer system to obtain information to support the assessment of the presence of Acute Respiratory Distress Syndrome (ARDS) in a person, comprising:

[0013] - a receiver unit,

[0014] - a control and a computing unit,

[0015] - an output unit, wherein the control and computing unit is configured to control the receiver unit to receive and / or to retrieve one or more radiologic image(s) of the chest of a person, wherein the control and computing unit is configured to process the one or more radiologic image(s) of the chest of the person by

[0016] 1) segmenting the one or more radiologic image(s) to identify different zones within the chest region of said one or more radiologic image(s), 2) applying a bone suppression algorithm to the one or more radiologic image(s) of the chest of the person to suppress bones in the one or more radiologic image(s),

[0017] 3) applying a feature extraction algorithm to the data obtained,

[0018] 4) calculating a probability of the presence of ARDS with the data obtained in step 3) by applying a classification and / or a regression algorithm, wherein the control and computing unit is configured to control the output unit to output the information of the calculation of a probability of the presence or ARDS according to step 4).

[0019] The present disclosure further provides a computer program product which when executed by a processor is configured to carry out the steps of: a) receiving and / or retrieving one or more radiologic image(s) of the chest of a person; b) processing the one or more radiologic image(s) of the chest of the person by: bl) segmenting the one or more radiologic image(s) to identify different zones within the chest region of said one or more radiologic image(s), b2) applying a bone suppression algorithm to the one or more radiologic image(s) of the chest of the person to suppress bones in the one or more radiologic image(s), b3) applying a feature extraction algorithm to the data obtained in steps a) and b); c) calculating a probability of the presence of ARDS with the data obtained in step c) by applying a classification and / or a regression algorithm; d) output the information obtained in step c).

[0020] The disclosure will be more particularly elucidated below without distinguishing between the subjects of the disclosure (method, computer system, computer program product). The elucidations that follow shall instead apply analogously to all subjects of the disclosure, regardless of the context in which they are made (method, computer system, computer program product).

[0021] Where steps are listed in a specific order in the present description, this does not necessarily mean that the steps must also be performed in the order specified. The disclosure shall instead be understood to mean that the steps listed in a specific order can be performed in any order or else in parallel with one another, unless one step is based on another step, which will be clear in each case from the description of the steps. The orders specifically listed in this document are therefore only preferred embodiments of the disclosure.

[0022] The present disclosure provides a computer-implemented method to obtain information to support the assessment of the presence of ARDS in a person. This shall assist physician and / or hospital staff to detect the onset of ARDS in a person who is typically an intensive care patient (a person who needs to be treated in an intensive care unit (ICU) of a hospital).

[0023] As a first step the computer-implemented method receives and / or retrieves one or more radiologic image(s) of the chest of a person (also referred to as step a)). This patient data (which may also include data of the person not relating to the one or more radiologic image(s) of the chest of the person) can be present, for example, in one or more data storage media that are part of a computer system or accessible via a network or via multiple networks. In particular, such data storage media can be one or more databases of a hospital, preferably databases in which patient data are stored. The patient data can be read from the one or more data storage media at defined times and / or at defined time intervals and / or upon occurrence of defined events (e.g., when new patient data become available) to use said patient data for the method according to the disclosure. It is also conceivable the patient data is retrieved / received from medical devices and / or computer systems connected to medical devices.

[0024] The term "medical devices" is understood to mean devices that can be used to obtain physiological information about a patient. Examples of such medical devices are heart rate monitors, blood pressure monitors, clinical thermometers, X-ray machines, computed tomography machines and the like. The medical devices are, in particular, X-ray machines, computed tomography machines, ventilators and devices which can be used to determine the value for partial pressure of arterial oxygen PaCh of a patient, the value for concentration of inspired oxygen Fi O2 and / or the PaOi / FiOi ratio and also values for PEEP and CPAP (PEEP = positive end-expiratory pressure; CPAP = continuous positive airway pressure).

[0025] The method according to the disclosure is as an example executed whenever new patient data are acquired and are stored in a data storage medium which can be part of the computer system according to the disclosure or to which the computer system according to the disclosure can be connected. The computer system according to the disclosure can be configured to check at defined time intervals (e.g., every 10 minutes) whether new patient data are present in the data storage medium. The one or more radiologic image(s) of the chest of a person are for example x-rays, dualenergy x-rays, CT scans (CT = computed tomography) and the like. Preference is given to x- rays image(s).

[0026] The radiological images are preferably present as digital images. The term “digital” means that the radiological images can be processed by a machine, generally a computer system. “Processing” is understood to mean the known methods for electronic data processing (EDP).

[0027] A "computer system" is an electronic data processing system that processes data by way of programmable computing rules. Such a system usually comprises a "computer", the unit that comprises a processor for performing logic operations, and also peripherals. In computer technology, “peripherals” refers to all devices that are connected to the computer and are used for control of the computer and / or as input and output devices. Examples thereof are monitor (screen), printer, scanner, mouse, keyboard, drives, camera, microphone, speaker, and the like. Internal ports and expansion cards are also regarded as peripherals in computer technology. Modern computer systems are frequently divided into desktop PCs, portable PCs, laptops, notebooks, netbooks, tablet PCs, handhelds (e.g., smartphones), cloud computers and workstations; all these systems can in principle be utilized for execution of the disclosure. Preferably, the computer-implemented method according to the present disclosure is executed on one or more computer systems in hospital intensive care units. It is conceivable for such a computer system to execute further functions in an intensive care unit, such as, for example, monitoring the state of health of intensive care patients. Preferably, the computer system can access databases of the hospital, in which patient data of the intensive care patients can be stored. Preferably, the disclosure is executed automatically on one or more computer systems as a background process. "Automatically" means that no human intervention at all is required.

[0028] Digital image files can be present in various formats. For example, digital image files can be coded as raster graphics. Raster graphics consist of a grid arrangement of so-called picture elements (pixel) or volume elements (voxel), to which a color or a gray value is assigned in each case. For the sake of simplicity, it is assumed in this description that the radiological images are present as raster graphics. However, this assumption is not to be understood as limiting in any way. There are a multitude of possible digital image formats and color codings; for a person skilled in the art of image processing, it is clear how the teaching in this description can be applied to different image formats. In step b) of the computer-implemented method according to the present disclosure the one or more radiologic image(s) of the chest of a person are processed as further discussed hereinafter. In step bl) (which is like step 1 of the computer system) the one or more radiologic image(s) are segmented to identify different zones within the chest region of said one or more radiologic image(s).

[0029] The segmentation step bl) in accordance with the present disclosure makes use of a trained machine learning model. In a preferred embodiment, the machine learning model is or comprises one or more convolutional neural networks (CNN). A CNN is a class of deep neural networks, most commonly applied to analyzing visual imagery (such as spectrogram representations). A CNN comprises an input layer with input neurons, an output layer with at least one output neuron, as well as multiple hidden layers between the input layer and the output layer. The hidden layers of a CNN typically consist of convolutional layers, ReLU (Rectified Linear Units) layer i.e., activation function, pooling layers, fully connected layers, and normalization layers. The nodes in the CNN input layer are organized into a set of "filters" (feature detectors), and the output of each set of filters is propagated to nodes in successive layers of the network. The computations for a CNN include applying the convolution mathematical operation to each filter to produce the output of that filter. Convolution is a specialized kind of mathematical operation performed by two functions to produce a third function that is a modified version of one of the two original functions. In convolutional network terminology, the first function to the convolution can be referred to as the input, while the second function can be referred to as the convolution kernel. The output may be referred to as the feature map. For example, the input to a convolution layer can be a multidimensional array of data that defines the various color components or grey scale values of an input image. The convolution kernel can be a multidimensional array of parameters, where the parameters are adapted by the training process for the neural network. More details about how to implement a convolutional neural network can be found in the literature (see e.g. Yu Han Liu: Feature Extraction and Image Recognition with Convolutional Neural Networks, 2018, J. Phys.: Conf. Ser. 1087 062032; H. H. Aghdam et al.: Guide to Convolutional Neural Networks, Springer 2017, ISBN: 978-3-319-57549-0; S. Khan et al.: Convolutional Neural Networks for Computer Vision, Morgan and Claypool Publishers, 2018, ISBN: 978-1-681-730219).

[0030] In a preferred embodiment of the present disclosure, the machine learning model for the segmentation step bl) is based on a specific kind of convolutional architecture called U-Net (see e.g. O. Ronneberger et al.: U-net: Convolutional networks for biomedical image segmentation, in: International Conference on Medical image computing and computer-assisted intervention, pp. 234-241, Springer, 2015, https: / / doi.org / 10.1007 / 978-3-319-24574-4_28 ). The U-Net architecture consists of two main blocks, an encoding path, and a decoding path. The encoding path uses convolutions, activation functions and pooling layers to extract image features, while the decoding path replaces the pooling layers with upsampling layers to project the extracted features back to pixel space, and finally recovers the image dimension at the end of the architecture. These are used in combination with activation functions and convolutions. Finally, the feature maps from the encoding paths can be concatenated to the feature maps in the decoding path in order to preserve fine details from the input data.

[0031] According to the disclosure the CNN and preferably the U-Net is configured and was trained on training data (radiologic images labeled by a clinical expert) to identify different zones within the chest region of the received one or more radiological image(s) of the chest of a person. For the training data the bone- suppressed images as obtained in step b2) can optionally also be used. Preferably, the trained machine learning model can segment following different zones within the image of the chest region of a person: lung quadrants: left upper, left lower, right upper, and right lower lung field, left and right hilus, heart, and the ribs (see e.g., also Fig 3). The image pre-processing can e.g. be done as described in Kai-Chih Pai et al., 2022, Digital Health, Vol. 8, 1- 15.

[0032] In step b2) of the method according to the present disclosure (which is like step 2 of the computer system) a bone suppression algorithm is applied to the one or more radiologic image(s) of the chest of the person to suppress bones in the one or more radiologic image(s).

[0033] Bone suppression algorithm include e.g. physical models such as a linear attenuation model which is used for dual-energy x-ray imaging. The linear attenuation model is a technique that can be used to separate the image information of different materials, such as bone and soft tissue, based on their varying attenuation properties at different X-ray energy levels. To generate an image where only the soft tissue is visible and the bone is suppressed, the process involves following steps: a) image registration (the high-energy and low-energy images are aligned to ensure precise correspondence between corresponding anatomical features, b) normalization (the pixel values in the high-energy and low-energy images are normalized to account for differences in acquisition conditions and detector response), c) a calculation of the linear attenuation coefficients (by using the pixel values of the high-energy and low-energy images, the linear attenuation coefficients of the materials at these energy levels can be calculated. These coefficients represent how much x-ray energy is absorbed as it passes through a material) and d) the generation of material- specific images (with the linear attenuation coefficients, a “bone image” and a “soft image” can be generated. Subtracting the bone image from the soft tissue image effectively suppresses the bone information, leaving only the soft tissue visible).

[0034] Another bone suppression algorithm includes the use of an inpainting algorithm and includes the steps of: a) subtracting a rib segmentation x-ray image (e.g. as obtained in step bl) from a full radiologic image of the chest of the person to generate a gap mask that encompasses areas outside the ribs, b) use of an inpainting algorithm such to reconstruct the missing information based on the processed image as obtained in step a). Inpaiting algorithm comprise neuronal networks or texture-based approaches. The neural networks can e.g., be generative models like Convolutional Neural Networks (CNNs) or Generative Adversarial Networks (GANs) and are trained on a dataset of images to learn the structure and features of images. A portion of these images contain gaps or missing areas. The trained networks can be used to predict the missing regions in the image. Texture -based approaches comprise e.g., texture-modeling algorithms where the model captures information about how the texture elements are distributed and connected on the image. This information is used to generate a coherent and consistent texture for the inpainted region. Texture -based approaches also comprise e.g., texture-transfer algorithms where the learned texture information from one part of an image is transferred to the inpainted region of the image. Commonly used texture -based algorithms are known and e.g., include patch-based approaches such as exemplar-based inpainting or patch-based textural inpainting, graph cut models such as graph cut texture synthesis or exemplar-based graph cut, Markov Random Fields (MRF) models such as texture fragment inpainting or poisson image editing.

[0035] In another preferred embodiment, a GAN is used as a bone suppression algorithm. As a preferred example of a GAN network architecture a Pix2Pix architecture can be utilized (Zhu et al., 2017, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 1125-1134). The GAN has a conditional framework an consists of two neural networks: 1) a generator (the generator takes the one or more radiologic image(s) and produces a bone-suppressed version of the one or more radiologic image(s)) and 2) a discriminator (the discriminator differentiates between real and generated images to determine whether the generated images are realistic or not. The discriminator also ensures that the generated images are consistent with the input images). For the training, a dataset of original and bone-suppressed images (e.g., manually annotated by clinical experts) and optionally the segmented images as obtained by step bl) are used to optimize the GAN. The training process involves a competition between the generator and the discriminator. The generator aims to generate images resembling real bone-suppressed images while the discriminator aims to differentiate between real and bone-suppressed images. As a result of the training, the generator becomes better at creating bone-suppressed images that match the desired characteristics. The dataset as used for the training process are dual energy x-ray images of the chest of persons. In a preferred embodiment, Haar 2D wavelet decomposed x-ray images of the chest of persons are used for the training process of the GAN (see Oh et al., 2018, Learning Bone Suppression from Dual Energy Chest X-rays using Adversarial Networks. CoRR (http: / / arxiv.Org / abs / l 811.02628)). In addition, the training preferably also involves history buffer and minibatch discrimination (see foresaid publication). After validation and optimization of the trained GAN the trained generator can be deployed to generate bone- suppressed images from one or more radiological images which are preferably x-ray images of the chest of a person. After the training of the GAN it is not required that dual energy chest x-ray images are used as input data. This is an advantage e.g., in comparison to the use of a physical model for bone suppression which requires the use of dual energy chest-x-ray images as input information.

[0036] In a preferred embodiment, steps bl) and b2) of the method of the present disclosure are combined and include the application of a trained machine-learning model. The trained machine-learning model is configured and was trained on training data to receive one or more radiologic image(s) of the chest of a patient and to identify different zones within the chest region of the received one or more image(s) and concurrently to generate bone suppressed images of the chest region of the received one or more image(s). The trained machine-learning model is a CNN or an image transformer. As indicated above, the trained machine learning model can preferably segment following different zones within the image of a chest region of a person: lung quadrants: left upper, left lower, right upper, and right lower lung field, left and right hilus, heart, and the ribs. At the same time, the CNN can generate bone suppressed images of the chest region of a person.

[0037] Besides a CNN also an image transformer technology can be utilized. Image transformer architecture comprises e.g., a multi-head self-attention module (which attends to the relationship between different parts of the image), an encoder-decoder architecture (the encoder divides the image into different hierarchical levels of features and extracts relevant information. The decoder uses this information to generate a segmented and bone suppression image output), positional encodings (are added to convey the spatial structure of the image to the model), and a segmentation head (to generate the segmentation and bone suppression image output). As with the CNN model the image transformer model is trained on a dataset of radiological images of the chest of persons and their corresponding segmentation masks and bone-suppressed images to be enabled to generate the intended output information. Image transformer architectures as used in the context of the present disclosure are e.g., discussed in following publication: Nicolas Carion et al. DETR: End-to-End Object Detection with Transformers 2020, ECCV.

[0038] In step b3) of the method according to the present disclosure (which is like step 3 of the computer system) a feature extraction algorithm is applied to the data as obtained in step bl) and step b2). Feature extraction in the context of step b3) involves the process of identifying an and extracting relevant patterns, textures, edges, and other distinctive attributes from the segmented and bone-suppressed one or more radiologic image(s) from the chest of a person. For this step b3) a region-based features analysis can be applied to calculate statistics and features directly from the segmented, bone-suppressed regions of the image(s). For instance, this relates to the computation of the area, perimeter, intensity mean and standard deviation etc. of each segmented object. In addition - and / or as an alternative - shape descriptors can be used to extract shape-related features from the segmented and bone- suppressed images which include aspects like ratios, circularity, or eccentricity. Another feature extraction approach applies texture analysis to capture subtle texture variations within segmented regions e.g., with algorithms such as gray-level-co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM) and / or local binary patterns (LBP) etc. An additional feature extraction approach includes the application of a contextual features approach which computes the spatial relationships or context with respect to other neighboring structures. All indicated feature extraction approaches can be applied individually. However, to obtain comprehensive and robust representation of image features all or some of these approaches are applied in parallel. In an alternative embodiment, the original one or more radiologic image(s) of the chest of a person can also be used as input information besides the segmented and bone-suppressed images.

[0039] In a preferred embodiment, the application of a feature extraction algorithm in step b3) includes patch sampling and image autoencoding.

[0040] In the patch sampling step the segmented and bone-suppressed images are divided into smaller patches. Features are now extracted from each patch using above discussed techniques such as e.g., a texture analysis (e.g., GECM). This step captures detailed information from different regions of the images. In the image autoencoding step an autoencoder is trained with the extracted patches. The autoencoder learns to encode the patches into a lower-dimensional latent space and decode them back to the original patch. The encoder part of the autoencoder is a feature extractor which maps each patch to a compressed representation in the latent space. The resulting encoded representations from the autoencoder serve as compact and meaningful feature vectors for each patch which capture the essence of the localized regions (while reducing redundancy and noise). The image autoencoder can e.g., be a convolutional autoencoder or a U-net.

[0041] In another preferred embodiment, the image autoencoder has a final classifier layer which is preferably an opacity classifier final layer. This opacity final classifier layer maps opacity features to classes (e.g. class 1: opaque and class 2: non opaque). The classifier is e.g., trained with a dataset of extracted features from the encoder (without the decoder part) and the corresponding labeled and classified data (labeling is done by a clinical expert). Opacity is one of the four Berlin criteria used to diagnose ARDS. If the one or more radiologic image(s) of the chest of a person show characteristic lung changes, such as opacity, one of the four Berlin criteria is fulfilled. The Berlin criteria will be discussed in detail below.

[0042] In step c) of the method according to the present disclosure (which is like step 4 of the computer system) a probability of the presence of ARDS with the data obtained in step b3) (equal to step 3 of the computer system) is calculated by applying a classification and / or a regression algorithm. In this calculation a value of 0 % can e.g., mean that ARDS can be ruled out and a value of 100% can e.g., mean that all evaluated patient data of the person indicate and / or infer that ARDS is present.

[0043] To obtain information to support the assessment of ARDS the Berlin Criteria is preferably utilized. The Berlin criteria comprise four main categories: 1) Timing: ARDS typically develops within a week of a trigger, such as a serious infection or a severe lung injury. An early onset is characteristic of ARDS; 2) Chest Imaging: occurrence of characteristic lung changes such as occurrence of bilateral opacities and fluid accumulation within the alveoli; 3) Etiology: the determination of the origin of the lung edema including the assessment whether a non- cardiogenic lung edema is present; 4) Oxygenation: measures by the ratio of the partial pressure of oxygen (PaOi) to the fraction of inspired oxygen ( FiOi) demonstrated reduced oxygen uptake (see The ARDS taskforce, 2012, JAMA, June 20, Vol 307, No. 23). The classification and / or regression algorithm used in step c) applies methods such as Support Vector Machines (SVM), Random Forest, K-nearest Neighbors (KNN). SVM seeks a separating line or hyperplane in the feature space that best separates between the presence of ARDS and the absence of ARDS. Random Forest aggregates the decisions of multiple individual decision trees, each trained on different subsets of data. This ensemble approach helps to mitigate overfitting and enhances the model's generalization capability. KNN works by finding the most similar data points in the training set to make predictions for new instances.

[0044] The input data for all approaches is the data obtained in step c) of the method according to the present disclosure. All approaches need to be trained with a dataset (the probability of ARDS for this dataset has been assessed by a clinical expert).

[0045] In another embodiment, the application of a classification and / or a regression algorithm in step c) includes the use of data of the person not relating to the one or more radiologic image(s) of the chest of the person. These patient data is preferably selected from the group of timing (onset in the past 7 days), origin of lung edema, oxygenation parameters (PaOi / FiOi). Therefore, this list includes the Berlin criteria 1), 3) and 4) as input data for step c) besides the chest image data that is processed according to steps a) and b). This information contributes to improve the calculation in step c) of the method according to the present disclosure.

[0046] In a preferred embodiment, the patient data further comprises clinical parameters and / or lab parameters as input data for step c) of the method according to the present disclosure. Clinical parameters comprise variables selected from the group of age, sex, body mass index, temperature, hear rate, systolic blood pressure, respiratory rate, vasopressor use at baseline, hematocrit, white cell count, bicarbonate, platelets, sodium, glucose, creatinine, bicarbonate, albumin, and bilirubin etc. These variables have already been used to identify ARDS (see Sinha P. et al., 2020, American Journal of Respiratory and Critical Care Medicine, Vol. 202, 7, Oct. 1). Lab parameters comprise variables selected from the group of Interleukin-8 (IL-8), Interleukin-6 (IL-6), plasminogen activator inhibitor- 1 (PALI), soluble tumor necrosis factor receptor- 1 (sTNFrl), surfactant protein D (SP-D), Willebrand Factor antigen (vWF), soluble intercellular adhesion molecule (sICAM-1), and protein C etc. (see Calfee C. et al., 2014, Lancet Respir. Med., 2(8): 611-620).

[0047] In another embodiment, in step c) of the method of the present disclosure (equal to step 4 of the computer system) the application of a classification and / or a regression algorithm also calculates an output data selected from the group of: probability of onset in past 7 days, occurrence of bilateral opacities in the one or more radiologic image(s), visualized opacity distribution on the one or more radiologic image(s), probability of a cardiogenic edema, prediction of oxygenation parameter (PaO2 / FiO2), prediction of mortality, prediction of ventilator-free days and prediction of ARDS sub -phenotype. The classification and / or a regression algorithm is trained with the corresponding dataset to enable the method to output one or more of the features discussed above.

[0048] In a preferred embodiment, in step c) of the method of the present disclosure (equal to step 4 of the computer system) the application of a classification and / or a regression algorithm also calculates an output data selected from the group of: probability of onset in past 7 days, occurrence of bilateral opacities in the one or more radiologic image(s), probability of a cardiogenic edema, prediction of oxygenation parameter (PaOi / FiOi). However, in this preferred scenario the indicated output data is calculated with the method according to the present disclosure with the one or more radiologic image(s) of the chest of a person as input data only. This has the advantage that one requires only the radiologic image data of the chest of a person to obtain information to support the assessment of the presence or ARDS according to all four main categories of the Berlin criteria (as all four main categories of the Berlin criteria are output data from the described method). This accelerates and simplifies the ARDS assessment. In a preferred embodiment, the output data obtainable with the method according to the present disclosure (based on one or more radiologic image(s) of the chest of a person as input data only) further comprises the prediction of mortality of the person, prediction of ventilator-free days (days where the person does not need ICU) and prediction of the ARDS sub-phenotype. The ARDS comprise e.g., two sub-phenotypes such as discussed in Calfee C. et al., 2014, Lancet Respir. Med., 2(8): 611-620. In this publication the hyperinflammatory sub phenotype (Phenotype 2) was characterized by higher plasma levels of inflammatory biomarkers, a higher prevalence of vasopressor use, lower serum bicarbonate, and a higher prevalence of sepsis, compared to Phenotype 1. Subjects in Phenotype 2 had higher mortality and fewer ventilator- free and organ failure-free days.

[0049] Another preferred subject matter of the present disclosure relates to the method as discussed herein wherein the output information in step d) includes at least the visualized opacity distribution on the one or more radiologic image(s) and the segmentation data obtained in step bl). This information provides a clinical expert such as a physician with optimal visual processed data to be enabled to identify the presence of ARDS and to make an independent judgement beyond the calculation of the probability of the presence of ARDS by the method according to the present disclosure.

[0050] Step d) of the method according to the present disclosure relates to the output of the data obtained in step c). The output can be an on-screen notification (e.g., a text message) and / or an audible signal and / or a visual signal and / or a vibrating alert and / or the like. The output can be output on one or more output units of the computer system according to the disclosure, for example displayed on a monitor and / or output via a speaker and / or output on a printer, for example as a text message, and / or transmitted, for example, via e-mail to a stationary computer system or a mobile receiving device (e.g., a smartphone or a tablet computer or a pager). The output is intended to support the assessment of ARDS of a person by a physician and / or hospital staff.

[0051] In another preferred embodiment, the output information in step d) is communicated as a message to the person and / or another person for further assessment of the case if the probability is above a defined threshold. In this scenario, the method according to the present disclosure is configured to determine an ARDS indicator value based on patient data. The ARDS indicator value correlates with the calculated probability of the presence of ARDS of the person. The ARDS indicator can have a value of 0 can mean that ARDS can be ruled out and a value of 1 or 100% can mean that all evaluated patient data indicate and / or infer that ARDS is present. The ARDS indicator value is compared with a defined threshold value. If there is a defined deviation between the ARDS indicator value and the threshold value, the method outputs a notification. Preferably, the method according to the disclosure is configured to output a notification if the ARDS indicator value is above the defined threshold value (the probability of the presence of ARDS in the person is high thus exceeding the threshold value). In such a case, measures preventing deterioration of the person’s condition should be taken. If the ARDS indicator value is below the defined threshold value, the probability of the presence of ARDS being present in the person is sufficiently low for no ARDS-specific measures to be necessary.

[0052] The threshold value can be, for example, a value between 0.5 (or 50%) and 1. Preferably, the threshold value is defined by a physician and / or based on medical experience, in particular on the basis of experience that can be gained by means of the present disclosure. For example, the threshold value can deliberately be set low to begin with to rule out any risk of overlooking ARDS in a person. The threshold value can then be adjusted over time to a higher threshold value that seems appropriate from a medical perspective and medical experience. Another embodiment of the invention relates to a computer system a receiver unit, a control and a computing unit and an output unit. All features as discussed herein for the computer- implemented method can also be implemented in the context of the computer system according to the present disclosure.

[0053] Another embodiment of the invention relates to a computer program product which when executed by a processor is configured to carry out the steps of the method according to the present disclosure. According to a further exemplary embodiment of the present invention, a computer readable medium, such as a CD-ROM, USB stick or the like, is presented wherein the computer readable medium has a computer program product stored on it which is / can be a computer program product as described herein. A computer program product may be stored and / or distributed on a suitable medium, such as an optical storage medium or a solid-state medium supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the internet or other wired or wireless telecommunication systems. However, the computer program product may also be presented over a network like the World Wide Web and can be downloaded into the working memory of a data processor from such a network. According to a further exemplary embodiment of the present invention, a medium for making a computer program product available for downloading is provided, which computer program product is arranged to perform a method according to the present disclosure.

[0054] The invention is described in detail hereinafter with reference to figures and examples, without wishing to restrict the invention to the features or combinations of features described and shown.

[0055] Figure 1 shows a schematic view of the method according to the present disclosure. The computer-implemented method 100 to support the assessment of the presence of ARDS in a person and comprises the steps of a) receiving and / or retrieving 110 one or more radiologic image(s) of the chest of a person; b) processing 120 the one or more radiologic image(s) of the chest of the person by: bl) segmenting 121 the one or more radiologic image(s) to identify different zones within the chest region of said one or more radiologic image(s), b2) applying 122 a bone suppression algorithm to the one or more radiologic image(s) of the chest of the person to suppress bones in the one or more radiologic image(s), b3) applying 123 a feature extraction algorithm to the data obtained in steps bl) and b2); c) calculating 130 a probability of the presence of ARDS with the data obtained in step b3) by applying a classification and / or a regression algorithm; d) output 140 the information obtained in step c). Figure 2 shows another schematic view of the method of the present disclosure. Optional features are shown in dotted lines. One or more radiologic image(s) of the chest of a person are used as input information 101. In step 120 the input data is processed by segmenting one or more radiologic image(s) of the chest of the person to identify different zones within the chest region. In step 122 the input data is processed by applying 122 a bone suppression algorithm to the one or more radiologic image(s) of the chest of the person. Depending on the model used data obtained in step 121 can optionally be used to train the algorithm of step 122 (and vice versa). Step 123 relates to the application of a feature extraction algorithm to the data obtained in step 121 and 122 (optionally including the original input data 101). The data obtained in step 123 is used as input data for the calculation of the probability of the presence or ARDS 130 which comprises the application of a classification and / or a regression algorithm. In an optional embodiment at least one of the features according of the Berlin criteria 1), 3) and 4) can be used in addition as input data for step 130. These features can be measured and assessed in the hospital with known methods and devices. Other additional optional input information for step 130 relate to clinical parameters 103 and / or lab parameters 104. The information obtained in step 130 is provided as an output in step 140 optionally together with other output information such as e.g., the original radiologic image 101, or a processed image according to one of the steps 121 and / or 122.

[0056] Figure 3 shows a schematic view of the segmentation step 121. The one or more radiologic image(s) of the chest of a person 101 is used as input information to generate segmented images 141 and 142. The image 141 shows following different zones within the image of the chest region of a person: lung quadrants: left upper lung field 1, left lower lung field 2, right upper lung field 3, right lower 4 lung field, heart 5, right hilus 6, and left hilus 7. The image 142 shows a chest image of the person with segmented ribs.

[0057] Figure 4 provides a schematic view of the bone-suppression step 122 and three different approaches possible to obtain the desired output information. Step 122a demonstrates a feasible way of implementing this bone-suppression step by a physical model such as a such as a linear attenuation model. The starting point of this method is a dual energy x ray image 101a of the chest of a person as input data. Step 122b shows another way of implementing bone-suppression application namely by using a GAN network architecture such as e.g., a Pix2Pix. After training of the GAN only x-ray images 101b of the chest of a person are required as input data. A third possible route of implementation 122c relates to the use of an inpaiting method which comprises a neuronal network or texture -based approach. This approach requires subtracting a rib segmentation x-ray image 142 (e.g., as obtained in step 121) from a full radiologic image of the chest of the person 101 and reconstructing the missing image information with an inpaiting algorithm. All approaches result in a bone-suppressed image as shown in 143 in Figure 4.

[0058] Figure 5 relates to a schematic view of the combined segmentation 121 and bone-suppression 122 approach by using the application of a trained machine-learning model. The trained machine-learning model is configured and was trained on training data to receive one or more radiologic image(s) of the chest of a patient 101 and to identify different zones within the chest region of the received one or more image(s) 141, 142 and to generate bone suppressed images of the chest region of the received one or more image(s) 143. The trained machine-learning model 121 & 122 is a CNN or an image transformer.

[0059] Figure 6 shows a schematic view of the feature extraction step 123. Independent of the feature extraction step 123 some (or all) of the so far obtained and / or processed image data (such as 101, 141, and 143) show radiomic features which are (or can) be useful for a physician to support the visual assessment of ARDS. This data can also be provided as output information 140. The exemplified embodiment of figure 6 shows a preferred way of implementing the feature extraction step 123 by using patch sampling 123a, an autoencoder 123b, and an opacity classifier final layer 123c. In the patch sampling step 123a the segmented images 141, 142 and the bone-suppressed images 143 are divided into smaller patches 144. The autoencoder learns to encode the patches 144 into a lower-dimensional latent space comprising latent features 123b-l and decode them back to the original patch 123b-2 (equals 144). The opacity classifier final layer 123c maps opacity last layer features 123c- 1 to classes (schematically depicted as class 0 or class 1 in Fig. 6).

[0060] Figure 7 shows a schematic view of the computer system 200 to obtain information to support the assessment of the presence of ARDS in a person, comprising a receiver unit 210, a control and a computing unit 220, an output unit 230. The control and computing unit 220 is configured to control the receiver unit 210 to receive and / or to retrieve one or more radiologic image(s) of the chest of a person. The computer system 200 is connected to a database 240 on which the computed one or more radiologic image(s) of the chest of a person are stored. It is also conceivable that the database 240 is part of the computer system 200. The control and computing unit 220 is configured to process the one or more radiologic image(s) of the chest of the person by segmenting the one or more radiologic image(s) to identify different zones within the chest region of said one or more radiologic image(s), applying a bone suppression algorithm to the one or more radiologic image(s) of the chest of the person to suppress bones in the one or more radiologic image(s), applying a feature extraction algorithm to the segmented and bone- suppressed images. A further step includes calculating a probability of the presence of ARDS with the data obtained in the feature extraction step by applying a classification and / or a regression algorithm. The control and computing unit 220 is configured to control the output unit 230 to output the information of the calculation of a probability of the presence or ARDS.

[0061] In the claims, the word “comprising” does not exclude other elements or steps, and the indefinite article “a” or “an” does not exclude a plurality. A single processor or other unit may fulfil the functions of several items re-cited in the claims. The mere fact that certain measures are re-cited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope.

Claims

Claims1. Computer-implemented method to obtain information to support the assessment of the presence of Acute Respiratory Distress Syndrome (ARDS) in a person, comprising the following steps:- a) receiving and / or retrieving one or more radiologic image(s) of the chest of a person;- b) processing the one or more radiologic image(s) of the chest of the person by: bl) segmenting the one or more radiologic image(s) to identify different zones within the chest region of said one or more radiologic image(s), b2) applying a bone suppression algorithm to the one or more radiologic image(s) of the chest of the person to suppress bones in the one or more radiologic image(s), b3) applying a feature extraction algorithm to the data obtained in steps bl) and b2);- c) calculating a probability of the presence of ARDS with the data obtained in step b3) by applying a classification and / or a regression algorithm;- d) output the information obtained in step c).

2. Method according to Claim 1, in which the processing steps bl) and b2) are combined and include the application of a trained machine-learning model, wherein the trained machinelearning model is configured and was trained on training data:- to receive one or more radiologic image(s) of the chest of a patient and- to identify different zones within the chest region of the received one or more image(s) and- to generate bone suppressed images of the chest region of the received one or more image(s).

3. Method according to Claim 2, wherein the machine-learning model is a neural network, preferably a Convolutional Neuronal Network (CNN) or an image transformer.

4. Method according to any one of Claims 1 to 3, wherein the application of a feature extraction algorithm in step b3) includes patch sampling and image autoencoding.

5. Method according to Claim 4, wherein the application of a feature extraction algorithm in step b3) also includes an opacity classifier final layer.

6. Method according to any one of the Claims 1 to 5, wherein the one or more radiologic image(s) are chest x-ray, dual energy x-ray and / or computed tomography image(s) of the chest of the person.

7. Method according to any one of the Claims 1 to 6, wherein the application of a classification and / or a regression algorithm in step c) includes the use of data of the person not relating to the one or more radiologic image(s) of the chest of the person.

8. Method according to Claim 7, wherein the data of the person not relating to the one or more radiologic image(s) of the chest of the person is selected from the group of timing (onset in the past 7 days), origin of lung edema, oxygenation parameters (PaOi / FiOi), clinical parameters and lab parameters.

9. Method according to any one of Claims 1 to 8, wherein in step c) the application of a classification and / or a regression algorithm also calculates an output data selected from the group of: probability of onset in past 7 days, occurrence of bilateral opacities in the one or more radiologic image(s), visualized opacity distribution on the one or more radiologic image(s), probability of a cardiogenic edema, prediction of oxygenation parameter (PaOi / FiCC), prediction of mortality, prediction of ventilator-free days and prediction of ARDS sub -phenotype.

10. Method according to any Claims 1 to 6, wherein, in step c) the application of a classification and / or a regression algorithm also calculates an output data selected from the group of probability of onset in past 7 days, occurrence of bilateral opacities in the one or more radiologic image(s), probability of a cardiogenic edema, prediction of oxygenation parameter (PaOi / FiCh).

11. Method according to Claim 10, wherein in step c) the application of a classification and / or a regression algorithm also calculates an output data selected from the group of prediction of mortality, prediction of ventilator-free days and prediction of ARDS sub-phenotype.

12. Method according to any one of the Claims 1 to 11 wherein the output information in step d) includes the visualized opacity distribution on the one or more radiologic image(s) and the segmentation data obtained in step bl).

13. Method according to any one of the Claims 1 to 12, wherein the output information in step d) is communicated as a message to the person and / or another person for further assessment of the case if the probability is above a defined threshold.

14. Computer system to obtain information to support the assessment of the presence of Acute Respiratory Distress Syndrome (ARDS) in a person, comprising:- a receiver unit,- a control and a computing unit,- an output unit, wherein the control and computing unit is configured to control the receiver unit to receive and / or to retrieve one or more radiologic image(s) of the chest of a person, wherein the control and computing unit is configured to process the one or more radiologic image(s) of the chest of the person by1) segmenting the one ore more radiologic image(s) to identify different zones within the chest region of said one or more radiologic image(s),2) applying a bone suppression algorithm to the one or more radiologic image(s) of the chest of the person to suppress bones in the one or more radiologic image(s),3) applying a feature extraction algorithm to the data obtained in step 1) and 2),4) calculating a probability of the presence of ARDS with the data obtained in step 3) by applying a classification and / or a regression algorithm, wherein the control and computing unit is configured to control the output unit to output the information of the calculation of a probability of the presence or ARDS according to step 4).

15. Computer program product which when executed by a processor is configured to carry out the steps of: a) receiving and / or retrieving one or more radiologic image(s) of the chest of a person; b) processing the one or more radiologic image(s) of the chest of the person by:bl) segmenting the one or more radiologic image(s) to identify different zones within the chest region of said one or more radiologic image(s), b2) applying a bone suppression algorithm to the one or more radiologic image(s) of the chest of the person to suppress bones in the one or more radiologic image(s), b3) applying a feature extraction algorithm to the data obtained in steps a) and b); c) calculating a probability of the presence of ARDS with the data obtained in step c) by applying a classification and / or a regression algorithm; d) output the information obtained in step c).

Images