Computer-implemented methods of quantifying and predicting progression of interstitial lung disease

Abstract

This paper discloses a computer-implemented method for quantifying and predicting the progression of interstitial lung disease. The method includes obtaining at least one image based on a scan of at least a portion of a patient's lungs, segmenting the image to obtain a lung mask defining the lung margin, identifying structures within the lung mask, and weighting the identified structures based on their relative positions to the lung margin to obtain a weighted score. The method can then include quantifying the degree of interstitial lung disease based on the weighted score, and / or predicting the progression of interstitial lung disease.

Description

Technical Field

[0001] This application relates to a computer-implemented method for quantifying the severity of interstitial lung disease and / or predicting its progression over time. Background Technology

[0002] Interstitial lung disease (ILD) is a group of conditions characterized by structural disorders of the lungs. In some cases, this can lead to irreversible and progressive fibrosis. Idiopathic pulmonary fibrosis (IPF) is a typical example of progressive fibrotic lung disease, characterized by a continuous decline in lung function and a life expectancy of 5 years after diagnosis. Other non-IPF fibrotic interstitial lung diseases also progress at a similar rate and have equally poor prognoses.

[0003] A standard technique used to measure lung function and the progression of ILD is the forced vital capacity (FVC) test, also known as a lung volume test. This involves measuring the amount of air an individual can forcefully exhale from their lungs after taking the deepest possible breath. The test may involve placing a special mask over the patient's face. The patient is then asked to inhale and exhale—as forcefully as possible—while data is collected.

[0004] Two antifibrotic drugs, nintedanib and pirfenidone, have reduced the rate of decline in lung function (FVC) in patients with IPF and non-IPFILD with progressive fibrosis. However, even in the case of IPF (the most homogeneous type of ILD), the disease trajectory varies among individuals to the point that prognostic information cannot be provided at diagnosis. Furthermore, the primary endpoint currently accepted in ILD clinical trials, variation in FVC, is plagued by at least 10% variability in measurement. Placebo-controlled trials in patients with IPF and ILD with progressive fibrosis are no longer ethically sound, thus demonstrating that continued improvement above standard of care presents a significant challenge to the pharmaceutical industry.

[0005] Therefore, there is an urgent need to discover and validate a new technology for identifying and predicting ILD in order to enrich clinical trials for patients more likely to worsen. This would shorten clinical trials, reduce sample sizes, and identify treatment responses above standard of care. Summary of the Invention

[0006] The various aspects of this application may be provided in combination with each other, and features of one aspect may be applied to other aspects.

[0007] In a first aspect, a computer-implemented method is provided for quantifying the severity of interstitial lung disease (optionally, and other solid lung lesions) and predicting its progression over time. The computer-implemented method includes obtaining at least one image based on a scan of at least a portion of a patient's lungs; segmenting the image to obtain a lung mask 200 defining the lung margins and identifying structures within the lung mask 200; and weighting the identified structures based on their relative positions to the lung margins to obtain a weighted score. Optionally, the method includes quantifying the current severity of the interstitial lung disease based on the weighted score and / or predicting its progression over time.

[0008] Advantageously, it has been found that computer-implemented methods for predicting the progression of interstitial lung disease (IFD) can identify not only patients whose disease is progressing or likely to progress, but also patients with stable lung function who are experiencing disease progression, addressing the challenge of FVC variability. Furthermore, by identifying patients whose disease is likely to progress, this method can be used to identify patients suitable for participation in new clinical trials. Moreover, computer-implemented methods can quantify the current degree of fibrotic disease (e.g., the current degree of ILD such as IPF and other lung parenchymal abnormalities) as well as existing indicators (such as pulmonary function or radiologist assessments), or may even outperform existing indicators (such as pulmonary function or radiologist assessments). For example, computer-implemented methods can include quantifying the current degree of ILD in terms of the volume or percentage of affected lung tissue.

[0009] Weighting identified structures based on their relative position to the lung edge to obtain a weighted score can include weighting only if the identified structure is within a selected threshold distance and / or a selected distance range from the lung edge. For example, the selected threshold distance could be, for instance, less than 1 to 2 cm, and / or the selected distance range could be, for instance, within 0.1 to 2 cm of the lung edge. For example, zero weight could be applied to regions identified as being outside the lung.

[0010] Segmenting an image to obtain a lung mask defining the lung margins may include determining the lung margins by identifying the lung-pleural boundary. The lung mask can be identified as a portion of the image of the normal lung parenchyma. Identifying the lung mask allows (1) the calculation of the lung margin repositioning relative to the lung margin.

[0011] (2) Quantify the total volume of the lungs so that the severity of the disease can be expressed as a percentage (e.g., disease severity volume / total lung volume = disease severity score).

[0012] Segmenting an image to identify structures within the lung mask may include identifying reticular vascular structures, which include at least one of blood vessels (e.g., pulmonary vessels) and reticular structures. Optionally, hypertransparent areas, such as airways, ground-glass opacities, emphysema, or cystic airways, may also be segmented. Segmenting an image to identify structures within the lung mask may additionally or optionally include identifying lung masses, which are then explicitly disregarded. Lung masses are considered features within the lung, but these areas are not considered to be associated with ILD. This means that any identified lung masses are included in the lung mask (and therefore in the calculation of total lung volume), but these masses are never considered reticular vascular structures or hypertransparent structures (and therefore are not included in the calculation of fibrotic disease volume; i.e., lung masses are always given zero weight).

[0013] The computer-implemented method may further include dividing at least one image into multiple distinct parts, such as pixels or voxels. And wherein weighting the identified structures based on their relative position to the lung edge to obtain a weighted score includes weighting each part of the image including the identified structures based on their relative position to the lung edge. For example, zero weight may be applied to features of an image that are not identified as lung tissue.

[0014] Predicting the progression of interstitial lung disease or quantifying the current severity of the disease based on the sum of weighted structural scores can include predicting the progression of interstitial lung disease based on the sum of weighted structural scores of all parts of an image.

[0015] In an example where each part of the image includes voxels, segmenting the image to obtain a lung mask defining the lung margins includes segmenting a 3D model to obtain a 3D lung mask defining the lung margins, and / or segmenting the image to identify structures within the lung mask may include obtaining normalized raw image data from a patient's lung scan and outputting a class probability for each voxel. The method may further include multiplying the sum of weighted structure scores by the volume of each voxel, and may further include dividing this value by the total number of lung voxels considered in the calculation to obtain a normalized quantification of disease severity as a percentage. Advantageously, this can yield a standardized score to account for acquisition resolution.

[0016] Obtaining at least one image based on a scan of at least a portion of a patient's lung may include obtaining multiple slices of the patient's lung. It should be understood that multiple slices can be obtained from a CT scanner. Each slice may represent a scan of a corresponding portion of the patient's lung. The method may also include creating a three-dimensional model based on the multiple slices of the patient's lung. Each slice preferably has a thickness of ≤3 mm, because if the slice thickness is too large, the structure of interest may not be visible in the scan.

[0017] Segmenting an image to obtain a lung mask defining the lung edges may include segmenting each slice to obtain a lung mask defining the lung edges for each slice. Segmenting an image to identify structures within the lung mask may include segmenting each slice to identify structures within the lung mask for each slice. A computer-implemented method may further include creating a three-dimensional model based on multiple slices of the patient's lung and providing the lung mask and the identified structures for each slice into the three-dimensional model. Weighting the identified structures based on their relative position to the lung edges to obtain a weighted score may include determining the relative position of the identified structures with respect to the edges of the three-dimensional lung mask.

[0018] Additionally or optionally, obtaining at least one image based on a scan of at least a portion of a patient's lung may include receiving a three-dimensional model of the patient's lung. In such an example, segmenting the image to obtain a lung mask defining the lung's edge may include segmenting the three-dimensional model to obtain a three-dimensional lung mask defining the lung's edge, and / or segmenting the image to identify structures within the lung mask may include segmenting the three-dimensional model to identify structures within the three-dimensional lung mask. Weighting the identified structures based on their relative position to the lung's edge to obtain a weighted score may include determining the relative position of the identified structures with respect to the edge of the three-dimensional lung mask.

[0019] Determining the relative position of the identified structure within the 3D model can include applying distance transformations, such as the Euclidean transform, to determine the distance from the lung edge of the 3D lung mask.

[0020] On the other hand, a computer-readable non-transitory storage medium is provided, including a program for a computer configured to cause a processor to perform the methods described above.

[0021] On the other hand, a method is provided for training machine learning models such as convolutional neural networks (CNNs) to identify intrapulmonary structures. This method may include using a dataset of whole-lung CT scans with corresponding labels for any one of four categories.

[0022] The machine learning model is trained in a manner where the four categories include background (corresponding to non-lung tissue), lung parenchyma (corresponding to normal, seemingly healthy lung tissue), pulmonary reticular vascular structures (corresponding to blood vessels, as well as other bright, elongated (reticular) structures present in fibrotic cases), and lung masses (corresponding to any abnormal, opaque, mass-like structures unrelated to clinical ILD). Optionally, a fifth category, including hypertransparent structures, may be present.

[0023] On the other hand, a device is provided for quantifying the severity of interstitial lung disease and predicting its progression, the device comprising:

[0024] A program module for obtaining at least one image based on a scan of at least a portion of a patient's lungs;

[0025] A program module for segmenting the image to (i) obtain a lung mask defining the lung edge, and (ii) identify structures within the lung mask;

[0026] A program module that weights the identified structures based on their relative positions to the lung margin to obtain weighted scores;

[0027] A program module that quantifies the severity of interstitial lung disease based on the weighted scores, and / or predicts the progression of interstitial lung disease over time.

[0028] In some implementations, the device may be a computer system.

[0029] In some implementations, the program module that weights the identified structures based on their relative position to the lung edge to obtain a weighted score is configured to only weight the identified structures if they are within a selected threshold distance from the lung edge.

[0030] In some implementations, a program module that segments the image to (i) obtain a lung mask defining the lung edge and (ii) identify structures within the lung mask is configured to determine the lung edge by identifying the pleural boundary and any lung masses within or attached to the pleural boundary.

[0031] In some implementations, the lung mask is identified as a portion of the image of a normal lung parenchyma.

[0032] In some implementations, the segmentation of the

[0033] The program module for (i) obtaining a lung mask defining the lung edge and (ii) identifying structures within the lung mask is configured to identify a reticular vascular structure, the reticular vascular structure including at least one of pulmonary vessels and a reticular structure.

[0034] In some implementations, a program module that weights the identified structure based on its relative position to the lung edge to obtain a weighted score is configured to divide at least one image into multiple distinct parts, and to weight each part of the image including the identified structure based on its relative position to the lung edge.

[0035] In some implementations, program modules that quantify the severity of interstitial lung disease based on the weighted scores and / or predict the progression of interstitial lung disease over time are configured to predict the progression of interstitial lung disease based on the sum of the weighted structural scores of all portions of the image.

[0036] In some implementations, a program module that obtains at least one image based on a scan of at least a portion of a patient's lung is configured to obtain multiple slices of the patient's lung.

[0037] In some implementations, a program module that obtains at least one image based on a scan of at least a portion of the patient's lungs is configured to create a three-dimensional model based on multiple slices of the patient's lungs.

[0038] In some implementations, a program module that segments the image to (i) obtain a lung mask defining the lung edge and (ii) identify structures within the lung mask is configured to segment each slice to identify structures within the lung mask for each slice; create a three-dimensional model based on multiple slices of the patient's lung; and provide the lung mask and the identified structures for each slice into the three-dimensional model.

[0039] In some implementations, a program module that weights the identified structures based on their relative position to the edge of the lung to obtain a weighted score is configured to determine the relative position of the identified structures relative to the edge of the three-dimensional lung mask.

[0040] In some implementations, a program module that weights the identified structures based on their relative position to the lung edge to obtain a weighted score is configured to apply a distance transformation to determine the distance from the lung edge of the three-dimensional lung mask. Attached Figure Description

[0041] Embodiments of this disclosure will now be described by way of example only, with reference to the accompanying drawings, in which:

[0042] Figure 1 A slice (cross section) of the patient's lungs from a CT scan is shown.

[0043] Figure 2 It shows Figure 1 The slices, in which a lung mask is applied to the scanned area to show the patient's lungs.

[0044] Figure 3 It shows Figure 1 The slices, in which a mask is applied to the scanned area to display the identified structure.

[0045] Figure 4A Another slice (cross section) of the patient's lungs from a CT scan is shown.

[0046] Figure 4B It shows Figure 4A The slices, in which a lung mask is applied to show the patient's lungs, including the scanned area containing lung masses.

[0047] Figure 4C It shows Figure 4AThe slices, where the mask is applied only to the reticular vascular feature.

[0048] Figure 5 A graph showing the probability of transplant-free survival for patients classified as low-risk or high-risk according to the FVC test, compared to the method of this disclosure;

[0049] Figure 6 A flowchart illustrating an example computer-implemented method for predicting the progression of interstitial lung disease is shown; and

[0050] Figure 7 A block diagram of a computer system suitable for implementing one or more embodiments of the present disclosure is shown. Detailed Implementation

[0051] Embodiments of this disclosure illustrate how to use baseline CT scans to quantify the current extent of fibrotic disease and predict outcomes for ILD patients (such as IPF), a method superior to FVC even when lung function parameters are controlled. Baseline CT scans can be analyzed to identify specific structures within the lung tissue. In particular, fibrosis-associated structures, such as blood vessels and thin (reticular) structures, are identified. Embodiments may also include identifying hypertransparent (low-density) structures within the lung associated with ILD, such as airways and cystic air spaces. Based on the presence and location of these identified structures, a trained machine learning model can be used to predict the severity and progression of ILD.

[0052] Advantageously, combining this imaging technique with FVC testing can be synergistic, as they can identify independent but complementary factors relevant to the outcome. In clinical trial settings, combining this imaging technique with physiology can improve patient selection and the definition of treatment response compared to traditional clinical physiological markers.

[0053] Figure 1 A slice (cross-section) of a patient's lung CT scan 100 is shown. This example image 100 was obtained from the Open Source Imaging Consortium (OSIC) database. Image 100 shows a cross-section of a patient's lung. Many structures within the lung tissue are visible. First, there is background non-lung tissue 103, which includes anything unrelated to the lungs, such as other organs and objects outside the body. Second, there is lung parenchyma 101, which may include normal, seemingly healthy, non-vascular lung tissue, including airways within the lungs. Third, there is pulmonary reticular vascular structure 102, which may include blood vessels as well as other bright, elongated (reticular) structures that appear in fibrotic cases. For example, reticular vascular structure 102 may include pulmonary vessels and reticular structures. Finally, there is a lung mass 105 (e.g., Figures 4A to 4CAs shown below (in more detail), these can be any abnormal, opaque, mass-like structure unrelated to clinical ILD. These may sometimes be tumors or incidental benign structures that do not resemble typical healthy tissue. In any case, they are unrelated to any fibrosis on the scan but may appear in the same scan within the lung. Alternatively, hypertransparent structures 107, such as airways or cystic air spaces, may also be identified as another type of structure.

[0054] Surprisingly, it was found that the progression of ILD could be predicted by identifying the presence and location of the pulmonary reticular vascular structure 102.

[0055] To be more detailed, Figure 1 The image 100 shown can be divided into multiple parts, such as pixels or voxels. Image analysis can be performed on image 100 by applying a machine learning model to obtain, for example... Figure 2 The lung mask 200 shown defines regions within the edges of lung structures (e.g., the edges of lung parenchyma 101, reticular vascular structures 102, and lung masses 105) and non-lung background regions (e.g., background non-lung tissue 103). Image analysis can also be applied to identify a second mask that identifies the reticular vascular structures 102 within the lung mask 200. Other features, such as lung masses 105, can be identified, allowing these features to be explicitly disregarded. This may also advantageously mean that any machine learning model is more robust and can produce reasonable outputs in the absence of fibrosis.

[0056] An example machine learning model for obtaining a lung mask 200 and identifying the reticular vascular structure could be a convolutional neural network (CNN), such as a 2D U-Net, which takes axial slices from a scan and outputs the segmentation of the slices. This U-Net model could include a variable number of block levels, such as four levels of blocks, containing a variable number (e.g., two) of convolutional layers with batch normalization and ReLU activation, as well as a max-pooling layer in the encoding part and an upper convolutional layer in the decoding part. The number of block levels and convolutional layers can be adjusted to change the performance of the architecture and can depend on the use case and can change with the training data. The number of convolutional filters in each block could be 32, 64, 128, and 256. The bottleneck layer could have 512 convolutional filters. Skip connections are used from the encoding layer to the corresponding layer in the decoding part. The input image could be a 2D CT scan of a patient's lung slices. The output could be a multi-channel probability map with anomalous regions of the same size as the input image. The channels can be converted into a binary segmentation mask by thresholding. Of course, any other segmentation architecture, including 3D architectures, could be used at this stage.

[0057] The identified reticular vascular structure 102 can then be weighted based on its relative position to the lung margin to obtain a weighted score. This weighted score can then be used to quantify the current extent of fibrotic disease and / or predict the progression of ILD.

[0058] Preferably, the weighting applied to the identified reticular vascular structure 102 is a function of its distance from the lung edge. In some examples, the weighting may be binary weighting. In such examples, weighting is applied only if the identified reticular vascular structure 102 is within a selected threshold distance from the lung edge, for example, less than 2 cm from the lung edge. In some examples, weighting is applied only if the identified reticular vascular structure 102 is within a selected threshold range from the lung edge, for example, between 0.1 cm and 2 cm from the lung edge.

[0059] It should be understood that, despite Figures 1 to 3 A two-dimensional scan of a patient's lungs is shown, but the method can be applied in three dimensions, for example, allowing weighting based on the three-dimensional nearest distance of the vascular network 102 to the lung edge. Two-dimensional (slice) or three-dimensional (whole volume) architectures can be used. Two-dimensional architectures can also employ multiple information channels, such as sliding windows of adjacent slices or maximum / minimum intensity projections of adjacent slices. This can provide the model with more environmental information about other slices surrounding the slice being cut, thus improving results. This can be done by expanding the input size of the convolutional neural network (CNN) architecture to accept images with multiple information channels (similar to how an RGB image contains three data channels corresponding to red-green-blue). In this example, the additional channels will have different features (e.g., the central slice in the first channel, the local maximum intensity projection in the second channel, and the local minimum intensity projection in the third channel). The CNN will need to be trained with these channels.

[0060] If this method is applied in 3D, the slices can be reconstructed together (this model has previously been applied to 2D slices to segment images to obtain a lung mask 200 and identify the reticular vascular structure within the lungs) to obtain a 3D binary lung mask 200 for the entire CT scan. This lung mask can then be used as input to a 3D Euclidean distance transform, which calculates the 3D shortest distance from each voxel to the lung edge. The output is then a 3D distance map that stores the distance of each voxel to the lung edge. Note that this is not a distance along the slice direction, but a volumetric distance.

[0061] Euclidean transform is a well-established image processing technique that uses a binary array (e.g., three-dimensional) and calculates the distance from each "1" voxel to the nearest "0" voxel by exhaustively examining the distance from each "1" voxel to each "0" voxel. Therefore, when the distance transform of the lung mask 200 is applied, the "1" voxel is the highlighted lung voxel (e.g., ...). Figure 2 (as shown), and calculates the distance of each voxel in the lung to the nearest non-lung "0" voxel. This distance takes into account the physical spacing of voxels in the scan, thus returning the distance in, for example, cm.

[0062] It should be understood that in some examples, other types of distance transformations (e.g., Manhattan) can be used, which are less precise but may be computed faster, so any type of distance transformation can be performed in this step depending on the desired accuracy of the obtained distance.

[0063] Alternatively, in some examples, instead of processing data piece by piece, if a 3D architecture is used, the 3D image data can be directly fed into the model, which can perform segmentation in 3D and then output the 3D segmentation.

[0064] Then, the distance graph is used as the weighting function.

[0065] Input. The weighting function can be binary (e.g., all identified reticular vascular voxels within a distance of X cm are given a weight of one, all other voxels are given zero weight and are not considered), or a continuous function, such as a linear or exponential function of the distance from the lung edge. Weighting can also consider the volume of each voxel, for example, by multiplying the weight by the volume of each voxel. Non-lung voxels can also be considered differently from voxels within the lung mask 200, for example, always given zero weight.

[0066] The optimal value for X has been found to be in the range of 0.1 to 2 cm, preferably in the range of 0.5 to 1.5 cm, and more preferably where X is 1 cm. Once X increases above 2 cm, the WRV (weighted reticular vascular) score becomes no longer more predictive than equal weighting of all RV (reticular vascular) voxels, regardless of location within the lung. The most effective point appears to be around 1 cm, and the further away from this value, the less reliable the weighted score becomes.

[0067] To obtain the final weighted score of the scan, the calculated weights of all identified reticular vascular voxels in the scan can be summed.

[0068] 1. The model (e.g., a 2D U-Net) can be a multi-class model trained to classify each voxel in the input array into one of four (or optionally five) mutually exclusive organization types:

[0069] 2. Background – Non-lung tissue. Anything unrelated to the lungs, such as other organs and objects outside the body.

[0070] 3. Lung parenchyma – normal, seemingly healthy non-vascular lung tissue, including the airways within the lungs.

[0071] 4. Pulmonary reticular vascular structure – blood vessels, and other bright, elongated (reticular) structures that appear in fibrotic cases.

[0072] 5. Lung masses – Any abnormal, opaque, mass-like structure unrelated to clinical ILD. These may sometimes be tumors, but are more likely to be incidental benign structures in the ILD population that do not resemble typical healthy tissue. Surprisingly, it has been found that identifying such lung masses, unrelated to any fibrosis on the scan, may appear in the same scan within the lung. These need to be detected so that they are not considered. This also means the model is more robust and can provide reasonable output even in the absence of fibrosis.

[0073] Alternatively, excessively light-transmitting structures—such as airways, emphysema, or cystic air spaces.

[0074] Therefore, the model can acquire normalized raw image data from a scan and output multi-channel class probabilities for each voxel. It then calculates which class of each voxel has the highest probability, providing a label map array (segmentation) of the same size as the model input, where the array values ​​specify the most likely tissue class for each voxel (i.e., 0 = background, 1 = solid, 2 = RV, 3 = mass / nodule, 4 = hypertransparent).

[0075] Then the label image is converted to the following lung mask 200 (e.g.) Figure 2 and Figure 4B (as shown) and reticular vascular mask (such as Figure 3 and Figure 4C As shown):

[0076] 6. Lung mask 200 (e.g.) Figure 2 and Figure 4B As shown): Construct a new binary mask in which any voxels to be segmented into lung parenchyma, reticular vascular region, hypertransparent region and lung mass are set to 1, and background voxels are set to 0.

[0077] 7. Mesh vascular mask 300 (e.g.) Figure 3 and Figure 4C As shown): Construct a new binary mask in which any voxel segmented into reticular vascular tissue is set to 1, and all other voxels are set to 0.

[0078] Optionally, any reticular vascular structures 102 identified outside the lung mask 200 can be automatically removed and treated as background non-lung tissue 103. The lung mask 200 can be defined by the presence of one of the following: solid 101, pulmonary reticular vascular structures 102, pulmonary hypertransparent structures 107, and pulmonary masses 105; therefore, by definition, such structures cannot occur outside the lung. In some examples, the method may include a post-processing step to clean the lung mask 200 and remove any false positive areas.

[0079] The model was trained on a dataset of whole-lung CT scans using the corresponding manually labeled data for the four categories mentioned above. This dataset is available from the Open Source Imaging Consortium (OSIC) database. The training set includes a combination of healthy scans and scans containing lung masses, allowing the model to learn what typical lung parenchyma and vessels look like and how to identify masses. All training scans can have an axial slice thickness of ≤3 mm, as the reticular vascular structure will not be visible in the scan if the slice thickness is too high. The dataset features various acquisition parameters, reconstruction kernels, and includes both contrast-enhanced and non-contrast-enhanced scans to provide a well-generalized training model.

[0080] The model was not trained on scans with fibrosis, although in some examples the model can be trained on scans labeled with fibrosis, and optionally, these can be segmented and used individually (optionally, exclusively) to obtain weighted scores for predicting the progression of interstitial lung disease. (See below for reference.) Figure 5 To describe it in more detail, even without specifically training the model using examples of fibrosis, the model already understands what a healthy lung looks like. This means that when the model sees a fibrotic reticular structure, it segments it into a reticular vascular structure because the reticular structure visually resembles blood vessels (in terms of strength and shape). Similarly, the model is able to identify blood vessels that have been stretched (tractioned) due to fibrosis because the stretched vessels are still clearly defined vascular structures (the edges of the vessels are more easily seen compared to healthy patients).

[0081] Figures 4A to 4C An example slice with a small mass (nodule) 105 is shown. Figures 1 to 3 Similar to the example, in Figure 4B In the middle, the lung shield 200 is covered in green. (For example...) Figure 4C As shown, the mask 300 for identifying the reticular vascular structure is covered in red. It can be seen how the lung mass 105 is contained within the lung mask 200, but not within the reticular vascular mask 300.

[0082] Figure 5A graph showing the probability of transplant-free survival for patients classified as low-risk or high-risk based on FVC testing, compared to the method disclosed herein (“e-ILD”). The above method was used to analyze CT scans of patients diagnosed with IPF or complicated by pulmonary fibrosis and emphysema (CPFE) using the OSIC database. Only those with full lung coverage and simultaneous FVC and carbon monoxide diffusion capacity (DLco) measurements (within 12 weeks post-acquisition) were included.

[0083] The model (“e-ILD”) takes as input 3D image data from a single CT scan, without any other clinical information. As described above, the model segments the lung-pleural boundary, identifies voxels that are part of the normal lung parenchyma, and voxels that include additional structures such as blood vessels, reticular formations, or hypertransparent areas (airways or cystic airways). The processed output includes a weighted score that quantifies the degree to which the lung is affected by reticular vascular abnormalities, giving the lung region a higher weight than other regions.

[0084] The analysis was limited to 278 patients (296 baseline CT series), who had:

[0085] 8. IPF or IPF / CPFE diagnosis;

[0086] 9. Effective baseline studies.

[0087] Patients with valid baselines and valid follow-up studies were included in the follow-up analysis (mean interval 54 weeks).

[0088] A study is considered valid if it meets the following criteria:

[0089] 10. Possess one or more original CT scans, fully covering the lungs, with slice thickness <= 3mm;

[0090] 11. Patients underwent FVC and DLCO measurements within 12 weeks of the study collection period.

[0091] Transplant-free survival was used as a surrogate measure of declining lung function in IPF patients as the primary endpoint of the analysis. Patients without vital sign data were reviewed at the final data point. Cox proportional hazards regression was used to model the relationship between lung function tests, identified reticular formation, and transplant-free survival. The predictive value of weighted scores based on identified reticular formation was assessed using Harrell's C index. Lung function parameters in the Cox regression model were adjusted to investigate the increase in weighted scores based on identified reticular formation. Imaging and lung function data were evaluated at baseline and follow-up. Relative changes from baseline to follow-up were also assessed, taking into account the time between the two points.

[0092] Baseline CT scans, concurrent pulmonary function tests, and outcome data were analyzed in 296 patients with IPF, including 120 patients with follow-up data. The mean follow-up interval was 54 (SD 13) weeks. At baseline, the C-index for FVC was 0.66. The C-index based on the weighted score of the identified reticular vascular structure was 0.75. The prognostic outcome of the e-ILD weighted score remained unchanged when FVC and DLco were further adjusted.

[0093] Patients were divided into low 403 and high 401e-ILD weighted groups (e.g. Figure 5 (As shown), and the low 405 and high 407 FVC groups, using the median as the threshold. In the high-weighted group, the observed risk of death was 4-fold higher (HR 4.0 CI 2.8–5.7, p<0.001). Patients categorized based on the median predicted FVC of 80% in the cohort showed a 50 / 50 prognostic outcome (HR 2.0, CI 1.4–2.8, p<0.001). When analyzing prognostic values ​​for relative change between baseline and follow-up, FVC reached a C-index of 0.56 for predicting survival from that point. However, by increasing the e-ILD weighted score, the C-index increased to 0.63. 31 patients (22%) had a relative FVC decrease of at least 10%. The e-ILD WRV score at baseline significantly predicted the likelihood of a decrease of at least 10% in FVC (OR 5.8 CI 1.5–22.3, p = 0.01, C-index 0.71).

[0094] Therefore, the data indicate that even when controlling for pulmonary function parameters, the weighted score of identified reticular vascular structures derived from baseline CT scans predicts IPF outcomes better than FVC. Combining changes in the weighted score based on identified reticular vascular structures and FVC between two time points outperforms changes in FVC alone in predicting outcomes, suggesting that imaging and pulmonary function testing identify independent and complementary factors relevant to outcomes. Furthermore, the e-ILD weighted score accurately predicts patients with a subsequent relative decrease in FVC of 10%.

[0095] Therefore, it has been demonstrated that the e-ILD weighted score predicts lung function decline and survival in IPF patients based on baseline CT scans. In clinical trial settings, combining automated imaging biomarkers with physiology improves patient selection and the definition of treatment response compared to traditional clinical physiological markers alone.

[0096] Figure 6 An example flowchart of a computer-implemented method for predicting the progression of interstitial lung disease is shown. Figure 6As shown, the method includes: acquiring 1010 at least one image comprising a scan of at least a portion of a patient's lungs; segmenting 1020 the image to (i) obtain a lung mask defining the lung margin, and (ii) identify structures within the lung; applying weights 1030 to the identified structures based on their relative positions to the lung margin to obtain weighted scores; and predicting the progression of interstitial lung disease based on the weighted scores 1040. Of course, it can be understood that in some examples, the step of predicting the progression of interstitial lung disease based on weighted scores 1040 may be optional, as it may be performed by different entities—for example, the model may simply output a weighted score or a quantification of the current disease severity, such as ml or %, and the raw scores or values ​​may be used by other parties.

[0097] Figure 7 This is a block diagram of a computer system 1200 applicable to implementing one or more embodiments of the present disclosure, including the methods implemented by the computer described above.

[0098] Computer system 1200 includes bus 1212 or other communication mechanisms for transmitting information data, signals, and information between various components of computer system 1200. These components include input / output (I / O) component 1204, which processes user (i.e., sender, receiver, service provider) actions, such as selecting keys from a keypad / keyboard, selecting one or more buttons or links, and sending corresponding signals to bus 1212. It may also include an optional camera or other means suitable for acquiring image data such as CT scans (although it should be understood that image data (e.g., CT scans) may be obtained elsewhere and transmitted to computer system 1200, for example, via network interface 1220).

[0099] I / O component 1204 may also include output components, such as display 1202 and cursor controller 1208 (e.g., keyboard, buttons, mouse, etc.). Display 1202 can be configured to present a user interface for viewing image data, such as... Figures 1 to 4CAs shown. Optional audio input / output components 1206 may also be included to allow a user to input information using speech by converting audio signals. Audio I / O component 1206 allows the user to hear audio. A transceiver or network interface 1220 sends and receives signals between computer system 1200 and other devices, such as another user device or a remote device that can provide database functionality via network 1222. In one example, the transmission is wireless, although other transmission media and methods may also be suitable. Processor 1214, which may be a microcontroller, digital signal processor (DSP), or other processing components, processes these various signals, for example, for display on computer system 1200 or for transmission to other devices via communication link 1224. Processor 1214 may also control the transmission of information to other devices, such as cached files (cookies) or IP addresses.

[0100] The computer system 1200 also includes system storage components 1210 (e.g., RAM), static storage components 1216 (e.g., ROM), and / or disk drives 1218 (e.g., solid-state drives, hard disk drives). The computer system 1200 performs specific operations by the processor 1214 and other components by executing one or more sequences of instructions contained in the system storage components 1210. For example, the processor 1214 can run applications 200 and 500 described above.

[0101] It will also be understood that the functionality of a computer program can be implemented in software or hardware, for example, as a dedicated circuit. For instance, it can be implemented as part of a computer system. The computer system may include buses or other communication mechanisms for transmitting information, data, signals, and messages between various components of the computer system. These components may include input / output (I / O) components that process user (i.e., sender, receiver, service provider) actions, such as selecting keys from a keypad / keyboard, selecting one or more buttons or links, and sending corresponding signals to bus 1212. I / O components may also include output components, such as displays and cursor controllers (e.g., keyboard, buttons, mouse, etc.). Transceivers or network interfaces can send and receive signals between the computer system and other devices, such as another user device or a remote server, via a network. In one implementation, the transmission is wireless, although other transmission media and methods may also be applicable. A processor (which may be a microcontroller, digital signal processor (DSP)) or other processing components processes these various signals, for example, for display on the computer system or for transmission to other devices via a communication link. The processor may also control the transmission of information to other devices, such as cached files or IP addresses.

[0102] Components of a computer system may also include system storage components (such as RAM), static storage components (such as ROM), and / or disk drives (such as solid-state drives, hard disk drives). A computer system performs specific operations by executing one or more sequences of instructions contained in the system storage components, which are then processed by the processor and other components.

[0103] Logic can be encoded in a computer-readable medium, which can refer to any medium that participates in providing instructions to a processor for execution. Such a medium can take many forms, including but not limited to non-volatile media, volatile media, and transmission media. In various embodiments, non-volatile media include optical discs or magnetic disks; volatile media include dynamic memory, such as system storage components; and transmission media include coaxial cables, copper wires, and optical fibers. In one embodiment, the logic is encoded in a non-transitory computer-readable medium. In one example, the transmission medium can be in the form of sound waves or light waves, such as sound waves or light waves generated during radio, optical, and infrared data communications.

[0104] Some common forms of computer-readable media include, for example, floppy disks, floppy disks, hard disks, magnetic tapes, any other magnetic media, CD-ROMs, any other optical media, punched cards, paper tapes, any other physical media with a perforated pattern, RAM, PROMs, EPROMs, FLASH-EPROMs, any other memory chips or cartridges, or any other media that a computer is adapted to read.

[0105] In various embodiments of this disclosure, a computer system may execute a sequence of instructions for practicing this disclosure. In various other embodiments of this disclosure, multiple computer systems 600 coupled to a network (e.g., such as a LAN, WLAN, PTSN, and / or various other wired or wireless networks, including telecommunications, mobile, and portable telephone networks) via communication links may execute sequences of instructions to practice this disclosure in a coordinated manner with each other.

[0106] It will also be understood that various aspects of this disclosure can be implemented using hardware, software, or a combination of hardware and software. Furthermore, where applicable, the various hardware and / or software components described herein can be combined into composite components comprising software, hardware, and / or both, without departing from the scope of this disclosure. Where applicable, the various hardware and / or software components described herein can be divided into sub-components comprising software, hardware, or both, without departing from the scope of this disclosure. Furthermore, where applicable, it is contemplated that a software component can be implemented as a hardware component, and vice versa.

[0107] The software according to this disclosure, such as program code and / or data, may be stored on one or more computer-readable media. It is also understood that the software identified herein may be implemented using one or more general-purpose or special-purpose computers and / or computer systems, networked and / or otherwise. Where applicable, the order of the various steps described herein may be changed, combined into compound steps, and / or divided into sub-steps to provide the features described herein.

[0108] The various features and steps described herein can be implemented as a system comprising one or more memories storing the various information described herein and one or more processors coupled to said one or more memories and networks, wherein the one or more processors are operable to perform the steps described herein as a non-transitory machine-readable medium comprising a plurality of machine-readable instructions, which, when executed by the one or more processors, is adapted to cause the one or more processors to perform a method comprising the steps described herein, and a method executed by one or more devices, such as the hardware processors, user equipment, servers and other devices described herein.

[0109] As can be understood from the above discussion, the embodiments shown in the accompanying drawings are merely exemplary and include generalizable, removable, or alternative embodiments as described herein.

[0110] Features of the replacement. Other examples and variations of the devices and methods described herein will be apparent to those skilled in the art within the context of this disclosure.

Claims

1. A computer-implemented method for quantifying the extent of interstitial lung disease and predicting its progression, the method comprising: obtaining at least one image based on a scan of at least a portion of a patient’s lung, wherein each of the at least one image comprises a plurality of pixels or voxels; segmenting the at least one image to obtain a lung mask defining a lung margin, the lung mask obtained by analyzing the at least one image via a trained machine learning model to determine margin structures, background regions, and reticular vascular structures; weighting each of the plurality of pixels or voxels associated with the determined reticular vascular structures based on their relative position to the determined margin structures to obtain a weighted score; quantifying the extent of interstitial lung disease and / or predicting the progression of interstitial lung disease over time based on the weighted score.

2. The computer-implemented method of claim 1, wherein the weighting each of the plurality of pixels or voxels associated with the determined reticular vascular structures based on their relative position to the determined margin structures to obtain a weighted score comprises weighting only when the identified structures are within a selected threshold distance of the lung margin.

3. The computer-implemented method of any of the preceding claims, wherein segmenting the image to obtain a lung mask defining a lung margin comprises determining the lung margin by identifying a lung pleura boundary and any lung masses within or attached to the lung pleura boundary.

4. The computer-implemented method of claim 1 or 2, wherein the lung mask identifies portions of the image that are normal lung parenchyma.

5. The computer-implemented method of claim 1 or 2, wherein segmenting the image to identify the reticular vascular structures comprises identifying the reticular vascular structures comprising at least one of lung blood vessels and reticular structures.

6. The computer-implemented method of claim 1, further comprising dividing the at least one image into a plurality of different portions; wherein, weighting the identified structures based on their relative position to the lung margin to obtain a weighted score comprises weighting each portion of the image comprising the identified structures based on their relative position to the lung margin.

7. The computer-implemented method of claim 6, wherein predicting the progression of interstitial lung disease based on a sum of weighted structure scores comprises predicting the progression of interstitial lung disease based on a sum of the weighted structure scores for all portions of the image.

8. The computer-implemented method of claim 7, further comprising multiplying the sum of the weighted structure scores by a volume of each voxel.

9. The computer-implemented method of claim 1 or 2, wherein obtaining at least one image comprising a scan of at least a portion of a patient’s lung comprises obtaining a plurality of slices of a patient’s lung.

10. The computer-implemented method of claim 9, further comprising creating a three-dimensional model based on a plurality of slices of a patient’s lung.

11. The computer-implemented method of claim 10, wherein: segmenting the image to obtain a lung mask defining the lung margin comprises segmenting the three-dimensional model to obtain a three-dimensional lung mask defining the lung margin; and segmenting the image to identify structures within the lung mask comprises segmenting the three-dimensional model to identify structures within the three-dimensional lung mask; and wherein weighting the identified structures based on their relative position to the lung margin to obtain a weighted score comprises determining the relative position of the identified structures to the edges of the three-dimensional lung mask.

12. The computer-implemented method of claim 11, wherein segmenting the images to obtain lung masks defining the lung margins comprises segmenting each slice to obtain a lung mask defining the lung margin of each slice; and segmenting the image to identify structures within the lung mask comprises segmenting each slice to identify structures within the lung mask of each slice; further comprising creating a three-dimensional model based on a plurality of slices of a patient’s lung, and providing the lung mask and the identified structures of each slice into the three-dimensional model; and wherein weighting the identified structures based on their relative position to the lung margin to obtain a weighted score comprises determining the relative position of the identified structures to the edges of the three-dimensional lung mask.

13. The computer-implemented method of claim 11 or 12, wherein determining the relative position of the identified structures within the three-dimensional model comprises applying a distance transform to determine a distance from the lung margin of the three-dimensional lung mask.

14. A computer-readable non-transitory storage medium comprising a program for a computer, the program configured to cause a processor to perform the method of any one of claims 1 to 13.

15. A computer system comprising a processor configured to: obtain at least one image based on a scan of at least a portion of a patient’s lung, wherein each of the at least one image comprises a plurality of pixels or voxels; segment the at least one image to obtain a lung mask defining a lung margin, the lung mask obtained by analyzing the at least one image via a trained machine learning model to determine edge structures, background regions, and reticular vascular structures; weight each of the plurality of pixels or voxels associated with the determined reticular vascular structures based on their relative position to the determined edge structures to obtain a weighted score; quantify a degree of interstitial lung disease based on the weighted score, and / or predict a progression of interstitial lung disease over time.

16. The computer system of claim 15, comprising a system storage component, wherein the system storage component is configured to contain one or more sequences of instructions for execution by the processor, wherein execution of the instructions by the processor is configured to cause the processor to: obtain at least one image based on a scan of at least a portion of a patient’s lung, wherein each of the at least one image comprises a plurality of pixels or voxels; segment the at least one image to obtain a lung mask defining a lung margin, the lung mask obtained by analyzing the at least one image via a trained machine learning model to determine edge structures, background regions, and reticular vascular structures.

17. The computer system of claim 15 or 16, comprising a camera configured to obtain at least one image of at least a portion of a patient’s lung, wherein each of the at least one image comprises a plurality of pixels or voxels.

Images