Assessment of lung disease progression
A machine learning approach is herein provided for preparing a model for assessing the progression of a lung disease, comprises receiving a first set of segmented images of lungs from different patients with the lung disease. The first set of images is segmented and used to train the model. The trained model is applied to a set of unsegmented images to generate a second set of segmented images. The model is updated with the second set of segmented segmentation. From the model, at least one result associated with progression of the lung disease is outputted.
1 . A computer-implemented method of preparing a lung lobe model for assessing progression of a lung disease, the method comprising:
receiving a first set of segmented images of lungs from patients with different severities of the lung disease, wherein the images in the first set are segmented using a trained lung segmentation model;
training the lung lobe model using the first set of segmented images based on segmentation of the images in the first set and lung morphology associated with the segmentation, wherein the lung morphology comprises at least lobes of left and right lungs;
applying the trained lung lobe model to a set of unsegmented images of lungs from different patients to generate a second set of segmented images, wherein the set of unsegmented images corresponds to the first set of segmented images;
updating the trained lung lobe model with the second set of segmented images; and
providing the updated lung lobe model configured to output at least one results associated with the lung disease progression of a patient.
2 . The method of claim 1 , wherein the first set of segmented images comprises segmented lung images masked by segmentation of the lungs derived using the trained lung segmentation model from a set of unsegmented computed tomography (CT) images of lungs.
3 . The method of claim 1 , wherein the trained lung segmentation model is configured to:
receive a set of unsegmented computed tomography (CT) images of lungs;
process unsegmented CT images of lungs at one or more resampling dimensions in relation to a probability map;
generate the probability map associated with said one or more resampling dimensions; and
provide the first set of segmented images from the unsegmented CT images of lungs based on the probability map.
4 . The method of claim 3 , wherein the first set of segmented images is provided at a full image resolution corresponding to an original image resolution of the set of unsegmented CT images of lungs.
5 . The method of claim 1 , furthering comprising: training a lung segmentation model, wherein the lung segmentation model is configured to:
receive a set of computed tomography (CT) images of lungs annotated with at least one designated segmentation;
process the set of annotated CT images of lungs based on the said at least one designated segmentation at one or more resampling dimensions;
generate the probability map associated with said one or more resampling dimensions in relation to said at least one designated segmentation; and
provide the trained lung segmentation model comprises the generated probability map.
6 . The method of claim 1 , further comprising: validating the lung lobe model with a new set of unsegmented images; and outputting the validation results from the lung lobe model.
7 . The method of claim 6 , further comprising: comparing the validation results to one or more annotated lung images of a healthy patient and/or a patient suffering from the lung disease to assess disease progression associated with each image from the set of unsegmented lung images.
8 . The method of claim 6 , further comprising: refining the lung segmentation model and/or lung lobe model based on an external input from a user, wherein the external input is associated with one or more corrections to the outputted validation results.
9 . The method of claim 8 , further comprising: iterating the refinement of the lung segmentation model and/or lung lobe model until each or both models achieve a pre-determined accuracy threshold in classifying one or more new sets of unsegmented lung images.
10 . The method of claim 1 , further comprising: outputting, from the trained lung lobe model, said at least one result associated with the lung disease progression of the patient based on a new set of unsegmented lung images from the patient.
11 . The method of claim 10 , wherein said at least one result is used to determine a dosage regimen for administering a drug to treat the lung disease.
12 . The method of claim 1 , wherein the lung disease is a type that affects lung airways.
13 . The method of claim 1 , wherein the lung disease is Idiopathic Pulmonary Fibrosis (IPF) or a related interstitial lung disease (ILD).
14 . The method of claim 11 , wherein said determining the dosage regime comprises achieving a clinical end-point for administering the drug for treating IPF or a related ILD.
15 . The method of claim 11 , wherein the drug is a type of antifibrotic.
16 . The method of claim 1 , wherein said at least one result is used for monitoring the progression of the lung disease in response to drug treatment.
17 . The method of claim 1 , wherein the outputted at least one result is used for improving clinical trial enrichment, wherein said at least one result is applied to select a patient with the lung disease at a desired disease progression.
18 . The method of claim 1 , wherein the lung segmentation model and/or lung lobe model is configured to capture the spatial and temporal dependencies in an input image.
19 . The method of claim 1 , wherein the lung segmentation model and/or lung lobe model is a type of convolutional neural network.
20 . A computer-implemented method of preparing an airway model for assessing progression of a lung disease, the method comprising:
receiving a first set of segmented images of lungs from different patients with the lung disease associated with airways, wherein the first set of images is segmented manually based on a generational level of airway;
training the model using the first set of segmented images, wherein the model is configured to segment the airways based on airway morphological parameters;
applying the model to a set of unsegmented images to generate a second set of segmented images, wherein the set of unsegmented images corresponds to the first set of segmented images;
updating the model with the second set of segmented segmentation; and
outputting, from the model, at least one result associated with the lung disease progression of a patient.
21 . The method of claim 20 , further comprising: validating the model with a new set of unsegmented images; and outputting the validation results from the model.
22 . The method of claim 21 , further comprising: refining the model based on an external input from a user, wherein the external input is associated with one or more corrections to the outputted validation results.
23 . The method of claim 22 , further comprising: iterating the refinement of the model until the model achieves a pre-determined accuracy threshold in classifying one or more new sets of unsegmented lung images.
24 . The method of claim 21 , further comprising: comparing the validation results to one or more annotated lung images of a healthy patient and/or a patient suffering from the lung disease to assess disease progression associated with each image from the set of unsegmented lung images.
25 . The method of claim 20 , wherein the airways comprise at least trachea, main bronchi, lobar bronchi, segmental and subsegmental bronchi.
26 . The method of claim 20 , wherein said segmented and unsegmented images comprise one or more types of computed tomography (CT) images of lungs.
27 . The method of claim 20 , wherein the first set of segmented images, the second set of segmented images, and the set of unsegmented images are generated using or derived from a device with an X-ray source.
28 . The method of claim 20 , wherein said airway morphological parameters comprise airway size, airway diameter, airway volume, pattern of airway branching, and degree of airway branching.
29 . The method of claim 28 , wherein the airway diameter is associated with a tissue condition in relation to a stage of disease progression, wherein the tissue condition comprises healthy issue and traction bronchiectasis.
30 . The method of claim 20 , wherein said at least one result is used for determining a dosage regimen for administering a drug to treat the lung disease.
31 . The method of claim 20 , wherein the lung disease is a type that affects lung airways.
32 . The method of claim 20 , wherein the lung disease is IPF or a related ILD.
33 . The method of claim 30 , wherein said determining the dosage regime comprises achieving a clinical end-point for administering the drug for treating IPF or a related ILD.
34 . The method of claim 20 , wherein said at least one result is used for monitoring the progression of the lung disease in response to drug treatment.
35 . The method of claim 30 , wherein the drug is a type of antifibrotic.
36 . The method of claim 20 , wherein the outputted at least one result is used for improving clinical trial enrichment, wherein said at least one result is applied to select a patient with the lung disease at a desired disease progression.
37 . The method of claim 20 , wherein the model is configured to capture the spatial and temporal dependencies in an input image.
38 . The method of claim 20 , wherein the model is a type of convolutional neural network.
39 . The method of claim 20 , wherein:
the first set of segmented images are from different patients with a lung disease that affects lung airways, including Idiopathic Pulmonary Fibrosis, IPF, or a related Interstitial Lung Disease, ILD;
the airway model comprises a multistage convolutional neural network configured to capture at least spatial dependencies in an input image;
the airway model is configured to segment, based on the captured spatial dependencies, the airways based on airway morphological parameters comprising airway size, airway diameter, airway volume, pattern of airway branching, and degree of airway branching;
the second set of segmented images are comparable to the first set of segmented images; and
said outputting comprises applying the updated airway model to one or more new unsegmented lung images from a patient obtained at different time points, to output at least one result associated with the IPF or ILD progression of the patient.
40 . The method of claim 39 , wherein the airway model is configured to receive a full-size three-dimensional CT image, wherein the full-size three-dimensional CT image is resampled to a plurality of different resolutions, wherein each of the plurality of different resolutions is input to a different stage of the multistage convolutional neural network, and wherein a probability map output from a previous stage at a first resolution is resampled and provided as input along with the resampled three-dimensional CT image at a second resolution, the first resolution being lower than the second resolution.
41 . The method of claim 40 , wherein the stages of the multistage convolutional neural network correspond to airway diameters in the airway morphological parameters.
42 . The method of claim 1 , wherein:
the first set of segmented images are from different patients with a lung disease including Idiopathic Pulmonary Fibrosis, IPF, or a related Interstitial Lung Disease, ILD;
the lung lobe model comprises a multistage convolutional neural network configured to capture at least spatial dependencies in an input image;
the lung lobe model is configured to segment based on the captured spatial dependencies;
the second set of segmented images are comparable to the first set of segmented images; and
said outputting comprises applying the updated lung lobe model to one or more new unsegmented lung images from a patient obtained at different time points, to output at least one result associated with the IPF or ILD progression of the patient.
43 . The method of claim 42 , wherein the lung lobe model is configured to receive a full-size three-dimensional CT image, wherein the full-size three-dimensional CT image is resampled to a plurality of different resolutions, wherein each of the plurality of different resolutions is input to a different stage of the multistage convolutional neural network, and wherein a probability map output from a previous stage at a first resolution is resampled and provided as input along with the resampled three-dimensional CT image at a second resolution, the first resolution being lower than the second resolution.