Method and system for detecting lung disease

Abstract

A system configured for use in diagnosing a respiratory health condition of a patient comprising: processor configured to: provide image data of lungs of a patient; derive two or more features from the image data; combine the two or more features to produce a composite feature; use the composite feature to diagnose the presence of a respiratory health condition in the lungs of the patient.

Description

[0001] METHOD AND SYSTEM FOR DETECTING LUNG DISEASE

[0002] The present application claims priority to Australian provisional patent application no. 2024901471 (filing date 19 May 2024), the entire content of which is incorporated herein by reference.

[0003] Background

[0004] In the past, X-ray imaging techniques such as Chest X-rays and CTs (Computed Tomography) have been extensively used in radiography to investigate internal organs, such as the lungs, in orderto visually diagnose disease. CTs are extensively used in medicine to provide a three- dimensional (3D) image of an area of interest, such as the lungs, which cannot easily be visually inspected. CTs are typically acquired using a dedicated x-ray machine that acquires data around a single axis of rotation, and this data is combined using computer processing (i.e. Computed Tomography) to produce tomographic images (virtual ‘slices’) of the lung. A physician (such as a radiologist) then visually inspects the CT slices to assess lung health, with their extensive training and experience allowing them to make a diagnosis.

[0005] CT can be used for imaging most body structures and has been particularly useful for detecting both acute and chronic structural changes inside the lungs.

[0006] However, many respiratory health conditions are difficult to diagnose using visual inspection of CT images. Patients experience the functional effects of the condition on their lungs, but the structural change cannot be seen in non-invasive images. This may be due to poor sensitivity of existing imaging methods and I or the microscopic scale of structural changes in the lungs, leading to the changes being unable to be detected by existing imaging techniques. These conditions are effectively invisible to current imaging techniques. This may be described as being unable to natively see the disease in the image.

[0007] Small airways disease is a feature of many of such respiratory health conditions that are difficult to detect using existing imaging techniques. In order to diagnose these types of respiratory health conditions, along with others, invasive surgery is required. Typically, this involves invasively extracting a piece of lung tissue for biopsy analysis. Preferably new diagnostic strategies for diagnosing respiratory health conditions are required. Summary

[0008] We have appreciated that certain diseases cannot “natively” be seen in images obtained by common imaging techniques. Embodiments extract functional measurements from the images (for example motion measurements using PIV) for use in diagnosing disease.

[0009] Embodiments of the invention aid the diagnosis of respiratory health disease from non- invasive images.

[0010] Embodiments aid the diagnosis of lung disease using biomarkers derived from image data (for example radiographic datasets) to detect the presence of lung disease.

[0011] Embodiments identify disease- and stage-specific XV biomarkers for COPD patients that correlated with airflow obstruction on spirometry. Further, we identified a set of XV-derived biomarkers that could distinguish veterans with DR-CB from controls despite normal spirometry in most patients from both groups.

[0012] Embodiments provide a safe and widely-available strategy for diagnosing respiratory health conditions disease while preserving spatial information.

[0013] In an embodiment the invention provides a method for use in diagnosing a respiratory health condition of a patient comprising the steps of: providing image data of lungs of a patient; deriving two or more features from the image data; combining the two or more features to produce a composite feature; using the composite feature to diagnose the presence of a respiratory health condition in the lungs of the patient.

[0014] In some embodiments, the steps of selecting a respiratory health condition for diagnosis, identifying a composite feature associated with the selected respiratory health condition, the composite feature comprising at least two features, wherein the derived two or more features at the at least two features comprised of the composite feature. Embodiments may comprise providing a separation function associated with the composite feature, the separation function, wherein the step of diagnosing the presence of the respiratory health condition is performed by comparing the composite feature with the separation function.

[0015] Embodiments may comprise a further step of calculating a biomarker, the biomarker being the comparison of the composite feature with the separation function. The biomarker may be an indication of the likelihood of the patient having the respiratory health condition. Embodiments may comprise the further step of deriving a dataset from the image data, wherein the two or more features are derived from the dataset. The step of extracting the value of the two or more features and the step of combining the features may be performed by combining the values of the features to produce a composite feature value, and use the composite feature value to diagnose the presence of a respiratory health condition in the lungs of the patient.

[0016] The feature may be a one dimensional representation of the image data. The features may be at least one of: functional features, defining a function of the lungs; and / or, structural features, defining structure of the lungs. The features may include temporal and spatial information. The dataset may be at least one of: a structural dataset; ventilation dataset; oscillation dataset; flow dataset; perfusion dataset.

[0017] The respiratory health condition may comprise: collateral ventilation disease or small airways disease. The features may provide information about functional factors causing the lung condition

[0018] In a further embodiment the invention provides a software application stored on a non- transitory medium adapted to perform the method of any one preceding embodiment.

[0019] In a further embodiment the invention provides a system configured for use in diagnosing a respiratory health condition of a patient comprising: processor configured to: provide image data of lungs of a patient; derive two or more features from the image data; combine the two or more features to produce a composite feature; use the composite feature to diagnose the presence of a respiratory health condition in the lungs of the patient.

[0020] In a further embodiment the invention provides a method for use in diagnosing a respiratory health condition of a patient comprising the steps of: acquiring image data of the lungs from at least two patients, wherein at least one of the at least two patients has a first lung condition and at least one other of the at least two patients does not have the lung condition; deriving two or more features from the image data for the at least two patients; combining the two or more features for each patient to produce a composite feature for each patient; generating a separating function, the separating function being a mathematical function of the derived two or more features which separates those patients of the at least two patients having the first lung condition from those patients of the at least two patients not having the condition; the separating function for use in diagnosing a respiratory health condition. The separating function may be generated according to the method of a previous embodiment.

[0021] In a further embodiment the invention provides a system for use in diagnosing a respiratory health condition of a patient comprising: Processor configured to: acquire image data of the lungs from at least two patients, wherein at least one of the at least two patients has a first lung condition and at least one other of the at least two patients does not have the lung condition; derive two or more features from the image data for the at least two patients; combine the two or more features for each patient to produce a composite feature for each patient; generate a separating function, the separating function being a mathematical function of the derived two or more features which separates those patients of the at least two patients having the first lung condition from those patients of the at least two patients not having the condition; the separating function for use in diagnosing a respiratory health condition.

[0022] A further embodiment of the invention provides a method of diagnosing lung conditions from a 4D dataset comprising the steps of: acquiring a 4D dataset of lung motion; selecting at least two features of the data set from a plurality of features; performing a data reduction process on the dataset to extract the at least two features from the 4D dataset; combining the extracted features; comparing the combination of extracted features to a predefined function associated with the combination of features and a lung condition to determine the likelihood of the presence of the lung condition.

[0023] The 4D dataset may be acquired from lung image data. For example, CT scans or fluorographs.

[0024] Examples of the features are described below. The predefined function may be a separating function separating combinations which exhibit a particular condition from those that don’t. The distance between the combination of extracted features and the separating function may be defined as a biomarker. The biomarker is used to define the probability of the patient having the condition or not.

[0025] Embodiments of the system may identify which combinations of features enable us to define a separating function between groups of patients with different lung conditions. And then, for a given patient (data set), compare the combination of features with the separating function to determine the likelihood of having the condition. A biomarker can be defined as the result of the comparison. Embodiments of the system identify combinations of features for which a reliable separating function can be defined between patients having a particular condition and those that don’t. Using a reliable combination / separating function, a patient’s 4D dataset can be acquired, the relevant features extracted and combined and then the combined features compares to the separating function to determine the likelihood of the patient having the condition.

[0026] The further the combination of extracted features is from the separating function, the greater the probability of the patient having the condition or not.

[0027] Brief Description of the Figures

[0028] In order that the invention be more clearly understood and put into practical effect, reference will now be made to preferred embodiments of an assembly in accordance with the present invention. The ensuing description is given by way of non-limitative example only and is with reference to the accompanying drawings, wherein:

[0029] Figure 1 shows an example of a CT machines.

[0030] Figure 2 shows a cross section of the scanning section of a CT machine.

[0031] Figure 3 shows an example slice of a CT.

[0032] Figure 4 shows a hypothetical example of using PIV to measure displacement.

[0033] Figure 5 shows a real world example of PIV showing a single coronal plane of lung tissue displacement measures.

[0034] Figure 6 shows an example of how perfusion can be calculated from the mass change in CT:VQ.

[0035] Figure 7 shows an example of extracting features from clinical data, applying a separating function, and assessing a patient for a condition.

[0036] Figure 8 shows an example of applying a separating function to different features.

[0037] Figure 9 shows an example of the successfulness of the separating functions shown in Figure 8.

[0038] Figure 10 shows an example of use of a separating function to aid in diagnosis.

[0039] Figure 11 shows an example of a separating function in a three dimensional feature combination.

[0040] Figure 12 shows an example of a method for aiding in diagnosis of a respiratory health condition.

[0041] Figure 13 shows an example of the use of a method for aiding in the diagnosis of a respiratory health condition.

[0042] Figure 14 shows an example of a system for implementing a method for aiding in the diagnosis of a respiratory health condition. Figure 15 shows an example of a method for aiding in diagnosis of a respiratory health condition.

[0043] Description

[0044] Embodiments derive information (a data set) from image data. The image data can for example be X-ray image data (e.g. fluoroscopy or CT), MRI image data, Nuclear Medicine image data (e.g. Nuclear VQ), etc. Acquisition of each of these types of image data are different, using different hardware and sometimes contrast agents, but they are all non- invasive, and therefore have the benefit of avoiding the risks, discomfort, and potential complications associated with surgical or needle biopsies, thereby providing a safer, quicker, and generally more comfortable diagnostic method for patients.

[0045] Acquiring lung image:

[0046] A) Fluoroscopy image acquisition

[0047] X-ray fluorographs (also referred to as cinefluorographs) contain 2D projections of 3D lung structure at multiple time points during the breath cycle, where the time elapsed between two frames is defined by the frame rate of the acquisition machine. To recover lung motion in 3D (for example using a technique such as XV LVAS, described later), X-ray fluorographs of the lungs are captured at multiple angles centred around the patient, with the fluorograph at each angle providing information on two out of the three spatial dimensions. These time series of X-ray images, in particular 2D X-ray images, can conceptually be considered as an X-ray movie of a patient breathing.

[0048] B) Computed Tomography (CT), paired-CT and 4DCT image acquisition

[0049] Computed Tomography (CT) is an advanced medical imaging technique used to obtain detailed cross-sectional images of internal structures within an object or body. Unlike traditional radiography, which projects x-rays through a subject onto a two-dimensional image detector, CT imaging employs a rotating X-ray source paired with detectors to capture multiple projections from numerous angles around the object being imaged. As the detectors measure the X-ray intensity after attenuation (absorption and deflection) by tissues and structures, the resulting projection data represents different degrees of density across the scanned object.

[0050] Referring to Figure 1 , a schematic diagram outlining the basic design of a CT system is shown. The CT system includes a circular CT scanner 20 including a rotating X-ray tube assembly and a detector array positioned diametrically opposite the X-ray tube assembly on the other side of the circle. The CT system includes an imaging bed 30 that is typically movable along the axis of the CT scanner. During scanning, a patient lies on the imaging bed 30 and is positioned within the CT scanner 20. When the patient is in position, the X-ray tube assembly and detector array rotate around the patient

[0051] Figure 2 shows a cross-sectional view through a CT scanner 200 with a patient 270 positioned within the scanner 200. An X-ray tube assembly includes X-ray tube 210, X-ray filter 220 and X-ray collimator 230. X-ray tube 210 is positioned within CT scanner 200 and rotates around the central axis of the CT scanner 200. X-ray filter 220 is positioned in front of the x-ray tube 210 and is tuned to absorb particular energy x-rays. X-ray tube collimator 230 aligns and directs the X-rays towards the area of interest of the patient. The X-ray beam is represented as 240. A detector array 250 is positioned diametrically opposite the X-ray tube assembly. The X-ray detector 250 detects X-ray radiation which has been attenuated by tissues of the patient’s body and converts it into a digital signal. The X-ray tube assembly and the detector array 250 are positioned within a cylindrical gantry 260. The imaging bed typically moves along the central axis of the gantry. The X-ray tube and the detector array spin within the gantry around the patient to capture X-ray images at different angles.

[0052] The acquired projection information undergoes computational processing known as image reconstruction. During this computational step, typically utilizing mathematical algorithms such as filtered back-projection or iterative reconstruction methods, the projection data is translated into a set of detailed two-dimensional image slices. These reconstructed cross-sectional images can be visualized individually or stacked to generate three-dimensional representations, accurately depicting internal anatomical and structural details with high spatial resolution. These slices show bones, blood vessels, soft tissues and other internal structures inside the body. An X-ray CT image contains 3D data of lung structure captured at a single time point during the breath cycle.

[0053] The dataset may be referred to as a 3-dimensional image dataset. The 3-dimensional image dataset may be represented visually. This visual representation is typically used by clinicians to help observe lung or other body part. The 3-dimensional dataset may be represented in a digital form, for example on a display screen of a digital device, for example a computer monitor, tablet screen, mobile phone screen etc. The 3-dimensional image dataset may be reproduced into a physical form. In the description, the terms CT image, image, 3-dimensional image, CT image dataset, or similar may be used to refer to the 3-dimension image dataset. Figure 3 shows an example slice (a coronal slice) of a CT image (typically doctors look at axial slices). The image shows a cross-sectional slice through the chest of a patient between the neck and the waist. A CT produces an image of the full 3-dimensional volume, the image of Figure 3 is a single slice through that 3-dimensional volume.

[0054] CT images are presented in black and white with varying shades of greyscale in between. These shades between black and white are referred to as the image intensity, but may also be referred to as shade, colour or brightness. The image intensity of a particular tissue feature in the image is determined by the degree to which it absorbs (attenuates) x-rays. Features of higher density have higher attenuation coefficients. The image intensity of a feature displayed in the image is determined by its mean attenuation. CT images use Hounsfield Units, and tissues with a high Hounsfield score have a high attenuation coefficient and appear white on the CT image (for example bones). Those features with a low Hounsfield score have a low attenuation coefficient and appear dark on the CT image (for example low density tissue, such as aerated lungs). CT scanners are calibrated so a particular density is always represented at the same image intensity. This allows calculations of density for the body parts within the image and comparison between images.

[0055] In the example image of Figure 3, the bone of the ribs is shown at 310 and appears white in the image due to its high attenuation coefficient. The body is shown at 320 and appears light grey, having a lower attenuation coefficient. Blood vessels 330 appear as light grey. The lung 340 appears in various degrees of dark grey having a low density. Airway 350 is the darkest feature of the image having the lowest density due to the high air content.

[0056] Standard CT imaging captures the lungs at a single respiratory phase, commonly during inspiration (full breath in). Alternatively, a paired inspiration-expiration (insp-exp) CT involves acquiring two sequential sets of CT scan images: one at full inspiration and another at full expiration (full breath out). Paired insp-exp CT imaging provides more information about the lungs, specifically by providing the shape of the lungs at more than one (i.e. two) different phase point, and thereby allow different datasets (such as motion and ventilation) to be extracted from the images that cannot be derived from a standard single-phase CT alone.

[0057] In addition to paired-CT, four-dimensional computed tomography (4DCT) provides the shape of the lungs at more than one different phase point (typically 7-10 phases, but between 3 and 20 phases). This is achieved by capturing sequential CT projections at various stages of inhalation and exhalation (generally during free breathing), and reconstructing these into a series of time-resolved, volumetric images (i.e. into separate CTs at each phase point). 4DCT also provides more information about the lungs by providing the shape of the lungs at more than one different phase point.

[0058] C) MRI and Nuclear VQ acquisition

[0059] In an alternative to X-ray and CT images, Magnetic Resonance Imaging (MRI) can provide detailed anatomical and functional images without exposure to ionizing radiation. MRI uses powerful magnetic fields and radiofrequency pulses to align and excite hydrogen atomic nuclei (protons), predominantly within the body's water molecules, and then detects the emitted signals as these protons relax back to their original states. By varying the parameters of image acquisition and leveraging differences in proton density, relaxation times, and molecular environments within tissues, MRI produces detailed contrast among soft tissues, making it particularly effective for imaging areas such as the brain, spinal cord, muscles, heart, and joints.

[0060] In a further alternative, the quantification of air flow (ventilation) and blood flow (perfusion) within lungs is commonly performed using Ventilation-Perfusion scintigraphy (nuclear VQ scans). These scans employ radiopharmaceuticals and are used in the diagnosis of a variety of lung pathologies including Pulmonary Embolism (PE), Pulmonary Hypertension (PH) and evaluation of lung transplants. Nuclear ventilation-perfusion (VQ) scanning is a process in which radioactive tracers are both inhaled (for the ventilation scan) and intravenously injected (for the perfusion scan), and the gamma rays they emit are detected via a gamma camera. These two separate scans allow indirect visualization of ventilation and perfusion, separately. Rather than producing an image of lung structure, nuclear medicine imaging directly provides lung ventilation and perfusion images.

[0061] The imaging techniques described above, including X-ray Computed Tomography (CT), X-ray fluorographs, and nuclear medicine imaging, provide image data of the lungs of a patient. Image data of the lungs can capture lung structure at different time points throughout the breath cycle, and different types of datasets can be derived from the image data.

[0062] Structural datasets, for example datasets containing structural information such as lung tissue types (normal, ground-glass opacity, consolidation, honeycombing, etc.) may be derived from the image data. Structural datasets define the lung structure. This may be defined at a single point in time. Structural datasets can be generated for different points of the breathing cycle. Structural datasets allow the structure of the lung to be analysed. Particular lung conditions may cause characteristic structural effects of the lung. 4D dataset of lung motion can be derived from the image data of the lungs. 4D datasets include spatial and temporal information. These datasets may be referred to as a spatiotemporal dataset. These 4D datasets may describe lung function and may be referred to as functional datasets.

[0063] 4D datasets of lung motion datasets are complex and can include 10,000 data points for each phase point of the breathing cycle. When data is captured at 14 phase points during the breathing cycle this creates 10,000 x 14 data points. Each data point represents a measurement of a local region of the lungs at a 3D spatial location (x, y, z) and at a time point (t) in the breath cycle.

[0064] Two examples of datasets, derived from image data, that describe lung function are ventilation datasets and perfusion datasets. These can be derived from the image data as described above.

[0065] Ventilation dataset can be a scalar field that describes the change in lung volume over time as the lung expands in volume and reduces in volume during the breathing cycle. This field is computed from a vector field, for example a 4D vector field, that represents lung motion throughout the breath cycle. This vector field is in turn computed by measuring the motion of lung structure at different time points in the breath cycle.

[0066] The perfusion dataset is also a scalar field, and it describes the flow of blood through the lungs, specifically within the pulmonary capillaries where gas exchange occurs between the blood and air in the alveoli. It is computed from a combination of a scalar field representing lung density and another scalar field representing the distance of a local region in the lung to the closest pulmonary vessel.

[0067] For a ventilation dataset, a small value represents a local region of the lung that has exhibited a small change in its volume throughout the breath cycle. Vice versa, a large value represents a local region of the lung that has exhibited a large change in its volume throughout the breath cycle. For a perfusion dataset, a small value represents low blood flow in a local region of the lungs. Vice versa, a large value represents high blood flow in a local region of the lungs. Each value provides a measurement of lung function in a local region of the lungs at one time point in the breath cycle. When represented as a 4-dimensional field, it provides a full spatial and temporally resolved measurement of lung function. Further datasets may be derived from the images. Also, further datasets may be derived to provide additional functional information about the lungs by transforming the datasets. These could be described as a second order transformation, or a second order dataset, since the image data has been transformed once to derive a first dataset (a first order dataset) and then that first dataset is transformed to derive a second dataset.

[0068] For example, image data may be transformed to derive a ventilation dataset to provide information about ventilation in the lung. That ventilation dataset may then be transformed to obtain functional information about flow in the lungs. For example, a flow dataset can be obtained by computing the difference in ventilation between two phase points divided by the time elapsed between two phase points. This flow dataset can provide information about air flow in the lungs.

[0069] In another example, transformation can be applied to the ventilation and perfusion datasets to obtain functional information about oscillation in the lungs. So an oscillation dataset can be derived.

[0070] As described above, many datasets can be derived from the lung image data to provide different information about lung function, including structural datasets, 4D datasets, spatiotemporal datasets, functional datasets, ventilation datasets, perfusion datasets, displacement datasets, velocity datasets, expansion datasets, heterogeneity datasets, flow datasets, oscillation datasets etc.

[0071] These datasets may be referred to as describing attributes of the lung. So the image data is transformed into a dataset which represents or describes that attribute of the lung, for example structure, ventilation, perfusion, flow, oscillation, etc.

[0072] Datasets derived from radiographic images and so may be referred to generally as radiographic datasets.

[0073] After retrieving image data of the lungs of the patient, a data set (also commonly written dataset) is derived from the image data. The method used to derive the dataset will depend on the image data that has been acquired, but also the desired dataset to be extracted. For example:

[0074] • If you have fluoroscopy images and you want to extract a ventilation dataset, such as a 4D ventilation dataset, then you could use a technique such as XV LVAS (also known as CTXV), which would provide a displacement (or motion) dataset from which you could further derive a ventilation dataset.

[0075] • If you have a paired-CT and want to extract a ventilation dataset, then you could use a technique such as CT LVAS, which would provide a displacement (or motion) dataset from which you could further derive a ventilation dataset.

[0076] • If you have a paired-CT and want to extract a perfusion dataset, then you could use a technique such as CTVQ (which would provide both a ventilation dataset and a perfusion dataset).

[0077] Furthermore, as a further brief example, regional lung ventilation data can be obtained using various techniques. Several techniques exist for obtaining regional ventilation measurements, and more specifically for non-invasively obtaining regional ventilation measurements. For example, Computed Tomographic X-ray Velocimetry (CTXV, also known as XV LVAS) exists for cinefluorograph sequences (a time series of 2D x-ray images), three-dimensional Particle Image Velocimetry (3DPIV) and Deformable Image Registration (DIR) exist for CT image data provides the shape of the lungs at more than one phase point, such as 4DCT sequences (a time series of 3D images) and paired inspiratory-expiratory CTs (two 3D images, one acquired at end-expiration and one acquired at end-inspiration), and analysis exists for nuclear medicine images (images of inhaled radioactive contrast agent, such as Xenon gas).

[0078] Some of these techniques will now be described. It will be understood that these are just a limited sample of methods for deriving a dataset from image data, and that other methods also exist.

[0079] A dataset, such as a four dimensional lung motion dataset describing the change in volume of the lung over time as the lung expands and reduces during breathing cycles, can be derived from images captured from fluorographs (cinefluorograph sequences, a time series of 2D x- ray images) during the breathing cycle using a technique called Computed Tomographic X- ray Velocimetry (CTXV, also known as XV LVAS).

[0080] CTXV (also referred to as XV LVAS commercially), which is often referred to as velocimetry techniques, measures the displacement of a region between two images, and then use the known time difference between the images to calculate the velocity of the region. It will be understood that, in certain circumstances, these techniques are often used to only measure the displacement of a region (without further calculating the velocity of the region). Patient lung data can be captured using fluoroscopy techniques (e.g. a number of cinefluorograph time sequences, on which XV LVAS style techniques can be used). This non- invasive technique provides in vivo internal detailed images of human organs.

[0081] Regional ventilation data is measured throughout all regions of the lung. Regional ventilation (V) is the volume of air that enters a particular lung region. It is expressed in a dimensionless form as specific ventilation, which is defined as the change in volume of a lung region between expiration and inspiration (AV), normalized by the volume of the same region at expiration (VO): V = (AV) / V0.

[0082] Regional specific ventilation can reveal abnormalities or changes to a patient’s breathing that are missed by global measurements (such as tidal volume, FEV1 , etc.). The measurement of regional specific ventilation is carried out by determining the expansion of thousands of discrete regions of the lung. In order to determine the expansion of each lung region the displacement of the lung is first calculated using a form of region tracking known as Particle Image Velocimetry (PIV). As shown in Figure 4, PIV involves discretizing a first image into multiple regions, identifying a region in the first image, locating that same region in the second image, and then calculating the displacement of that region between those images. CTXV (XV LVAS) extends the PIV technique furtherto utilise the cinefluorograph sequences (time series) from multiple angles (typically four or five angles or “views”) in order to reconstruct a 3D displacement field for each phase point, without first reconstructing a 3D image. Each data point represents a measurement of lung function in a local region of the lungs at a 3D spatial location (x, y, z) and at a time point (t) in the breath cycle. Having a 3D displacement field for each phase point creates a 4D displacement dataset.

[0083] Once the displacement vector field is known for all regions in the lung, the regional displacement vectors can be used to calculate the regional expansion of the lung tissue. It will be understood that these are in fact two datasets, a displacement dataset and an expansion dataset. It can be assumed that the majority of the lung volume change of any given region is occurring due to air entering the lungs, and thus the expansion dataset can be consider (or further modified to become) a ventilation dataset.

[0084] A dataset, such as a four-dimensional lung motion dataset describing the change in volume of the lung overtime as the lung expands and reduces during breathing cycles, can be derived from images captured from a paired-CT using a technique called Three-dimensional Particle Image Velocimetry (3DPIV, or CT LVAS when commercially used to measure lung ventilation). 3DPIV (also referred to as CT LVAS commercially), which is often referred to as velocimetry technique, measure the displacement of a region between two 3D images, and then use the known time difference between the images to calculate the velocity of the region. It will be understood that, in certain circumstances, these techniques are often used to only measure the displacement of a region (without further calculating the velocity of the region).

[0085] Patient lung data can be captured using CT imaging (e.g. a paired-CT or a 4DCT, on which 3DPIV or deformable image registration techniques can be used). This non-invasive technique provides in vivo internal detailed images of human organs.

[0086] Similar to XV LVAS, 3DPIV provides a 4D ventilation dataset with regional ventilation data is measured throughout all regions of the lung, and can also provide regional specific ventilation. 3DPIV uses a 3D version of traditional 2D PIV to measure the displacement of 3D regions of the CT. 2D PIV involves discretizing a first image into multiple regions, identifying a region in the first image, locating that same region in the second image, and then calculating the displacement of that region between those images. Figure 4 shows an example of using 2D PIV to measure displacement by: (1) splitting the image into multiple regions and selecting a region as shown in Image 1 ; (2) finding that same region in the second image as shown in Image 2; and (3) calculating the displacement based on the distance that the region has travelled as shown in Image 3. For 3DPIV this process is carried out on 3D regions.

[0087] For 3DPIV this process is then repeated for each region in the CT, resulting in a 3D vector field that describes the 3D displacement of the lung tissue from the expiration CT to the inspiration CT. Figure 5 shows an example of PIV showing a single coronal plane. Image 1 shows a Pre-processed Expiration Image; Image 2 shows a Pre-processed Inspiration Image (in which the lungs appear larger); and, Image 3 shows a PIV displacement field (with the vectors showing the displacement measurements for the lung tissue between expiration and inspiration for the CT slice shown). Having a 3D displacement field for each phase point creates a 4D displacement dataset.

[0088] Ventilation and perfusion data for the lung may be acquired from paired CT images using a technique referred to as CT:VQ Analysis Software (CT:VQ). This software-based image processing technology analyzes two non-contrast thoracic Computed Tomography (CT) reconstructed volume scans (one at inspiratory phase and one at expiratory phase) to quantify airflow (ventilation, V) and blood flow (perfusion, Q) in the lungs. CVTQ outputs regional ventilation data and regional perfusion data. In CTQV, the device first calculates regional ventilation (V) by measuring the regional displacement of lung tissue between inspiration and expiration (from the paired inspiration / expiration CT images). The device performs this by measuring the motion of the lung tissue at thousands of points in the lung using three-dimensional Particle Image Velocimetry (PIV), an established imageprocessing technique (as described above in relation to CT LVAS, and shown in Figure 4 and Figure 5). Alternatively, deformable image registration (DIR), or any other suitable regional displacement measurement technique, could be used in place of PIV. The device then calculates regional ventilation from this motion (displacement) field based on the local volumetric expansion of each three-dimensional element between the two points in time (i.e. between the inspiration CT and the expiration CT), a well-established and accurate measure of the net increase (or decrease) of air in the element over that time period. Having a 3D displacement field for each phase point creates a 4D displacement dataset.

[0089] CTVQ further calculates regional perfusion (Q) by calculating the mass change, or relative density (Hounsfield unit values), between the inspiration and expiration CTs for each region of lung tissue (with the displacement measurements being used to ensure that the same region of lung tissue, which is a different volume in the two CTs due to the increased volume caused by breathing, is assessed). The difference between the mass in the inspiration CT and the expiration CT is accounted for by the change in blood volume in the lungs, which CTVQ characterizes as lung perfusion (blood flow).

[0090] Calculation of the regional perfusion data is measured throughout all segments of the lung. Perfusion of the lungs refers to the process of transporting blood to the pulmonary capillaries. The device measures regional perfusion by quantifying the change in mass of a lung region between expiration and inspiration (AQ). Regional perfusion can reveal abnormalities or changes to a patient’s pulmonary vascular function that can indicate lung pathophysiology.

[0091] Measuring regional perfusion can be carried out by determining the mass change of thousands of discrete regions of the lung. In order to determine the mass change of each lung region the following steps are carried out:

[0092] 1 . Discretizing the expiration CT into multiple regions, and selecting a region;

[0093] 2. Locating that same region in the inspiration CT (by using the information from the PIV displacement field calculated in the regional ventilation steps described in the previous section);

[0094] 3. Calculating the mass of the region in the expiration CT,

[0095] 4. Calculating the mass of the region in the inspiration CT, and 5. Determining the change in mass of the region between the expiration CT and the inspiration CT.

[0096] The paired inspiration-expiration CT allows for the direct calculation of mass change due to CTs using calibrated Hounsfield Units (HU). This is carried out by converting both the expiration CT and the inspiration CT from HU to mass, and then subtracting the expiration mass from the inspiration mass to calculate the mass change.

[0097] Figure 6 shows an example of how the mass change is calculated by: (1) splitting the expiration CT into multiple regions and selecting a region (bold box), as shown in Figure 3 Image 1 ; and, (2) locating that same (expanded) region in the inspiration CT as shown in Figure 3 Image 2. The mass of the region is then calculated in both the expiration CT and the inspiration CT, and the change in mass of the region is calculated. Having a 3D mass change field for the patient creates a 3D perfusion dataset.

[0098] Datasets, in particular large or complex datasets (e.g. 4D datasets), can be described in terms of ‘features’. Describing a dataset using a feature can simplify the data in the dataset. Many different features can be used to describe a dataset, and in some cases thousands of different features can be used to describe a dataset. A feature may also be referred to as a metric, and deriving a feature can often involve mathematical equations to extract the desired information. It will be understood that these features will often be a mathematical relationship derived from one or more properties of the dataset. In this way, features act to simplify the dataset, and more specifically features act to reduce the dimensionality of the dataset.

[0099] A feature is a 1-dimensional representation of the data points of the dataset. For datasets involving multiple phase points in a breath, a feature may be defined at each phase point of the breathing cycle (e.g. airflow, with the airflow being provided for each phase of the respiratory cycle). Given 4-dimensional datasets describe lung function, feature extraction can be performed on these datasets.

[0100] A feature may be derived by combining or aggregating a dataset to distil the dataset down to a 1 -dimensional representation. For 4-dimensional datasets, as these 4-dimensional datasets are spatially and temporally resolved, these 1 -dimensional features can be described in different, broad categories of features such as spatial features, temporal features, and spatial- temporal features (sometimes called spatio-temporal features). The process of deriving the feature from the dataset may be described using terms including: extracting, simplifying, reducing, combining, aggregating, resolving, distilling. Spatial features represent a feature in which a subset of all the data points in the lungs that can be grouped together spatially in a physiologically meaningful way, such as grouping the data points into the left and right lung regions, or grouping the data points into its primary lobes. In this way a spatial feature is one that takes into account regional variation (e.g. it isn’t just global lung measurements). Temporal features represent a feature in which a subset of all the data points in the lungs that can be grouped together temporally, such as grouping the data points at the phase points representing the start to the middle of the breath cycle. In this way a temporal feature is one that takes into account temporal variation (e.g. it isn’t just total ventilation for a region of lung tissue). Spatial-temporal features aggregate both spatial and temporal information into a 1 -dimensional representation.

[0101] Different types of features exist, where different feature types describe a different function in different terms. Some examples of feature types are described below, including flow features, divergence features, Oscillatory Velocity Index (OVI) features, Shannon entropy features, etc. This list is not comprehensive but provided just as an example.

[0102] A few examples of features of a 4D lung motion data set are now described. These examples are not a comprehensive list of features but provided as examples.

[0103] In the following descriptions of features there are seven measurements across the breath (i.e. there are seven phase points of data), these are referred to as “t=0-6” etc. The first measurement is t=0, then t=1 , up to t=6. Reference to t=2-4 means that it is only using the middle 3 measurements for calculating the features (it isn’t using t=0, t=1 , t=5, t=6).

[0104] OVI Features:

[0105] Oscillatory Velocity Index (OVI): This group of features describes lung function as the amount of oscillation (or volatility) in the regional motion of the lungs over different phase points during breathing (in particular tidal inspiration). These features are extracted from a displacement (or motion) dataset, and are typically extracted from a 4D displacement dataset. Given a dataset (vector field p) representing lung motion at different phase points, OVI can be defined as Where dt represents the time elapsed between two phase points as a proportion of the tidal inspiration breath. Two examples of OVI features are:

[0106] 1. OVI Skewness: As OVI is measured at a regional level of the lung, there is a measurement for each region of the lung, and these multiple measurements can be further assessed as a distribution of OVI measurements over the whole lung. OVI skewness is a feature that measures the degree of asymmetry in this distribution. This degree of asymmetry can be computed using the Fisher-Pearson coefficient of skewness, which is the ratio of the 3rd central moment over the standard deviation of the distribution cubed.

[0107] 2. OVI Kurtosis: Similar to the OVI skewness, OVI kurtosis is a feature that describes the distribution of OVI measurements over the whole lung. In contrast to OVI skewness, OVI kurtosis measures the shape and spread of the distribution relative to a normal distribution. More specifically, this feature describes the degree of “tailedness”, or the amount of “mass” present in the tails of the distribution relative to a normal distribution. The “tailedness” can be computed as a ratio of the 4th central moment over the standard deviation of the distribution to the power of 4.

[0108] Flow Features:

[0109] This group of features describes lung function by measuring the regional air flow through the lungs during tidal inspiration. Flow features can be derived from ventilation datasets (ventilation datasets can be derived from displacement datasets, for example by computing the expansion from the regional displacement dataset). Airflow is defined as the regional rate of lung ventilation, and can be defined as where t denotes the phase point, At represents the time elapsed between two phase points in seconds, and V_t represents the ventilation measured at time t. Some examples of flow features are:

[0110] 3. Sum of mean x IQR of flow (t=1-4): Similar to the oscillation-based features, regional measurements of air flow form a distribution over the whole lung. This feature describes the spread / heterogeneity of air flow multiplied by the mean air flow aggregated overthe middle phases of tidal inspiration. The spread / heterogeneity of air flow is computed using the interquartile range, which is the difference between the 75th percentile and 25th percentile of the distribution. This is a spatial-temporal feature.

[0111] 4. Sum of mean flow (t=1 -4): Similar to (3), this feature describes the distribution of regional air flow measurements over the whole lung. In contrast to (3), at every phase point during tidal inspiration, mean air flow is measured. These measurements are then aggregated in the middle phases of tidal inspiration.

[0112] 5. Sum of high region of flow (t=2-4): Similar to (3), this feature describes the distribution of regional air flow measurements over the whole lung. This feature describes air flow in a spatiotemporal manner. It measures the proportion of lung volume that has regional air flow above a certain pre-defined value at every time point. These measurements are then aggregated throughout the middle phases of tidal inspiration.

[0113] 6. Maximum flow skewness (t=2-4): Similar to (3), this feature describes the distribution of regional air flow measurements over the whole lung. At every phase point, the degree of asymmetry in the air flow distribution is measured. This feature then measures the maximum degree of asymmetry in air flow distribution throughout the middle phases of tidal inspiration.

[0114] 7. Maximum flow inverse defect percentage (t=0-6): Similar to (3), this feature describes the distribution of regional air flow measurements over the whole lung. At every phase point, the proportion of lung volume with a value of regional air flow greater than a certain scale factor of the mean air flow is computed. The maximum of these values is then measured by this feature throughout tidal inspiration.

[0115] Divergence Features

[0116] This group of features is derived from a ventilation dataset. Divergence features estimate lung function by measuring the amount of regional specific ventilation in the lungs throughout tidal inspiration. Given a Jacobian J that describes the spatial derivatives of lung motion from start inspiration, regional expansion is defined as where det(*) represents the determinant.

[0117] 8. Divergence Mean (t=6): Using the distribution of regional specific ventilation over the whole lung at the end of tidal inspiration, this feature measures the average specific ventilation over the whole lung.

[0118] 9. Divergence IQR (t=6): Using the distribution of regional specific ventilation over the whole lung at the end of tidal inspiration, this feature describes the spread of the regional specific ventilation distribution. The spread of the distribution is computed using the interquartile range of the distribution, which is the difference between the 75th percentile and the 25th percentile of the distribution.

[0119] 10. Divergence Skewness (t=6): Using the distribution of regional specific ventilation over the whole lung at the end of tidal inspiration, this feature measures the degree of asymmetry in average specific ventilation distribution over the whole lung. This degree of asymmetry is computed using the Fisher-Pearson coefficient of skewness, which is the ratio of the 3rd central moment over the standard deviation of the distribution cubed.

[0120] In order to reduce the complexity of the 4D dataset, which has thousands of regional measurements for each phase point, and multiple phase points across the breath, a process can be used to extract features. This process is referred to as feature extraction, or sometimes data reduction.

[0121] A dataset is first selected. The dataset may be selected based on an attribute, for example a displacement dataset, velocity dataset or flow dataset may be selected. It will be understood that some datasets can be derived from other datasets, for example a velocity dataset can be derived from a displacement if the time over which the displacement occurred (e.g. the time between two images that are used in PIV analysis to measure motion) is known. A subattribute may also be selected. For example, in a velocity dataset, maximum velocity, average velocity or modal velocity may be selected as a sub-attribute. If a sub attribute is selected then the dataset is transformed to a dataset for the sub attribute (for example if the maximum velocity is selected as the sub-attribute, then the velocity dataset would be transformed into a maximum velocity dataset). The selected dataset for the desired attribute is combined in space and time to distil the dataset down to a single number, referred to as a feature. The process of distilling may be referred to as ‘reducing’.

[0122] The process of reducing the dataset down to a feature can involve computing a discrete distribution, for example computing a discrete distribution for a spatial region of the dataset, or for example computing a discrete distribution for a spatial region of the dataset for each phase point of the breathing cycle. Using these discrete distributions at different spatial regions and at each phase point, an attribute is used to represent each discrete distribution, with examples such as the maximum value, the average value, the modal value of the distribution, and so on. These attributes can then be combined / aggregated in space and time to distil the 4-dimensional field down to a 1 -dimensional representation, referred to as a feature. In this way, the step of deriving a feature from a dataset involves reducing the dimensionality of the dataset, and in particular reducing the multidimensional dataset down to a 1 -dimensional feature.

[0123] This process of reducing the dataset to a feature can involve using all of the spatial and temporal values in the dataset, or a subset of either the spatial, the temporal, or both the spatial and temporal measurements in the dataset. For example, the entire spatial information could be used, but in a case where there are seven phase points, only three of the seven are used for calculating the feature (for example the middle three phase points). In other words, the data reduction can use either the full dataset, or a subset (i.e. a subset in the spatial (x,y,z) or temporal (t) domains) to extract the feature.

[0124] A simple example of the feature extraction process is now provided. Using the ventilation field (dataset) as the attribute of interest, and assuming that the spatial region of interest is the entire lung region, a discrete distribution of the ventilation field is computed at each phase point of the field. Assuming that the attribute used to represent each discrete distribution is the average, the average of the ventilation field at the final phase point of the breath cycle can be used as the extracted 1 -dimensional feature representation of the field.

[0125] For each patient, for each set of images, two or more features can be extracted from the various datasets and combined to produce a composite feature (sometimes called a combined feature). The features are combined to provide a multidimensional combination. Since there are many features for each lung dataset, there are many potential combinations of features for a lung dataset. There are also multiple ways that features may be combined, for example using different algorithms. The number of combinations can be extremely large, for example hundreds of different features, or thousands of different features, or 10,000 different features, may be extracted from a single dataset.

[0126] Some combinations of features correlate to provide a strong indication of the likelihood of the presence of a lung condition. Other combinations do not. In other words, for some combinations there is a clear variation (or clear distinction, or clear separation) between patients having a particular condition and those patients that do not have the particular condition. In some cases a function may be defined, sometimes called a separation function, which defines values which separate those patients having a particular condition and those that do not have the particular condition. It will be understood that such a separation function does not need to be 100% accurate in the separation, but that instead it only needs to be clinically useful for the detection of disease (i.e. it provides appropriate specificity and sensitivity).

[0127] It will be understood that the features for determining whether or not the patient has a particular health condition need to be known in advance. The composite feature may be determined through clinical trial(s) or other suitable methods. For example, and referring to Figure 7, a clinical trial may be run that has two patient cohorts, one with the lung condition and one without (i.e. a control group). Figure 7a shows the results of the clinical trial investigating Condition A, where two features (Feature B and Feature C) are being investigated to determine whether these features correlate to provide a strong indication of the likelihood of the presence of a lung condition. In Figure 7a each patient in the trial is represented by a symbol on the plot, with their value for Feature B and value for Feature C defining their position on the plot. Patients with Condition A are represented by a plus (+) and patients without Condition A are represented by a circle (O). For example, one patient 720, who is part of the control group, has a Feature B value of 0.5 and a Feature C value of 0.9, while another patient 730, who is part of the Condition A group, has a Feature B value of 0.9 and a Feature C value of 3.4.

[0128] Figure 7b visually shows the separation function 710 that can mathematically be applied to the clinical trial data in order to divide the patients into two cohorts, in this instance with a strong indication of the likelihood of the presence of a lung condition (this will not be the case for all combinations of features). It will be noted that separating function 710 incorrectly classifies two of the patients in the group (711 , 712). Patient 711 is incorrectly classified as having Condition A (a false positive), and Patient 712 is incorrectly classified as not having Condition A (a false negative). However, so long as appropriate specificity and sensitivity are achieved, this is considered acceptable. Some conditions (diseases) may have particular physical factors that make one type of feature more appropriate for detecting the condition. For example, some conditions may require having at least one spatio-temporal feature, and some may require at least one spatio-temporal feature derived from a 4D dataset, and some may require at least one spatio-temporal feature derived from a 4D dataset having at least 3 phase points.

[0129] Figure 7c provides an explanation of how extraction of Feature B and Feature C can be helpful to future patients. For example, Patient 715 sees as doctor and has medical images taken of their lungs, and from the image data a value for Feature B and a value for Feature C are determined (0.8 for Feature B, and 3.7 for Feature C), and this composite feature value is used to determine whether Patient 715 has Condition A. This process is shown visually in Figure 7c, where Patient 715 is represented by a square, and has been plotted with their value for Feature B and Feature C (0.8; 3.7). As can be seen, Patient 715 is in the Condition A region of the plot, thus indicating that they likely have Condition A.

[0130] By identifying combinations that provide a distinct variation between patients having a specific lung condition and those that do not, the system can use the composite feature to determine the likelihood of whether a patient has that specific lung condition or not. It will be understood that different conditions will likely have different combinations of features that will provide the best separation function. In addition, it will also be understood that a single condition will possible have multiple different combinations of functions that will work to provide a good separation function.

[0131] In an example, in which the likelihood of a patient having lung condition A can be assessed by a composite feature comprising Feature B and Feature C, the system can:

[0132] • retrieve image data of the lungs of a patient,

[0133] • extract Feature B and Feature C,

[0134] • combine Feature B and Feature C to produce the composite feature, and

[0135] • use the composite feature to detect the presence of Lung Condition A.

[0136] In another example in which the likelihood of a patient having lung condition A can be assessed by a composite feature comprising Feature B and Feature C, the system can:

[0137] • retrieve image data of the lungs of a patient, • derive the required dataset (or datasets) for Feature B and Feature C from the image data,

[0138] • extract the value of Feature B and the value of Feature C from the dataset(s),

[0139] • combine the value of Feature B and the value of Feature C to produce the composite feature value, and

[0140] • use the composite feature value to detect the presence of Lung Condition A.

[0141] Embodiments enable the presence of lung disease to be detected from image data, where the lung disease is not visible in the images and has not previously been detectable by known non-invasive techniques.

[0142] The features that are computed from the datasets can be carefully selected so that the extracted features provide a physiologically meaningful representation of lung function. In other words, the features can be selected so that they attempt to measure the spatial and temporal heterogeneity of lung function, with the core assumption that the distribution of healthy lung function is homogeneous. Once the list of all possible features has been narrowed down to the more physiologically meaningful features, the composite feature selection process enumerates through all possible combinations of these features to identify combinations of features that provide a strong indication of the likelihood of the presence of a lung condition. This process can be carried out for combinations of 2 features (a 2D combination), combinations of 3 features (a 3D combination), combinations of 4 features (a 4D combination), and so on. Assuming that the features that have been selected provide complementary information of lung function, typically combinations of features from different fields of interest work well. Combinations of different features from a single field typically work less well because they are describing the same distribution, and hence tend to provide less useful information when combined together. In other words, the combination of one feature from a ventilation dataset and one feature from a flow dataset may provide a better separating function than two features from a velocity dataset. As such, having a composite feature that includes a first feature from a first dataset, and a second feature from a second dataset, where the first dataset and the second dataset are different, can be helpful.

[0143] Figures 8a and 8b provide examples of combinations of features. In each graph a different composite feature is calculated. The composite features are presented in graphical form for the purposes of illustration. In some embodiments, the composite feature may be presented graphically, for example to a clinician, to enable the clinician to make a visual assessment of whether or not a patient has a particular lung condition. In other embodiments, the assessment may be performed automatically based on the value of the composite feature. Figure 8a provides an example where the 2D combination (Feature B and Feature C) allows for good separation of the data by a separation function 810, and Figure 8b provides an example where the 2D combination (Feature X and Feature Y) do not allow for good separation of the data by a separation function (attempted separating function 820).

[0144] Referring to Figure 8a, a function (sometimes referred to as a separating function) 810 is generated for each combination of features. Separating function 810 is represented here as a dashed line. Separating function describes a mathematical function which attempts to best separate one group of patients and another (for example a group of patients with a lung condition, for example COPD, and a group of patients without the lung condition (control patients)). Separating function can be defined mathematically in terms of the values of the two features. Depending on the combinations of features selected, a function may or may not exist which can fully separate the two groups (e.g. Figure 8b). For some combinations of features a function may exist for a particular lung condition that separates patients having that lung condition from those that do not have that condition. For other combinations of features, a function may not exist for a particular lung condition that separates patients having that particular lung condition from those patients not having that particular lung condition.

[0145] Functions that separate the groups of patients for a particular lung condition can be used to provide an indication of whether a patient can be classified into one group or another. Such functions can be used to determine the presence of a specific lung condition by extracting the relevant features, combining the features to produce a composite feature, and comparing the composite feature to the function to determine whether the patient has the specific lung condition.

[0146] Some functions provide a strong correlation (i.e. have a strong probability of distinguishing one group from another) and others do not show a strong correlation (i.e. have a low probability of distinguishing one group from another and so would be unreliable). These combinations which produce a function having a strong correlation are useful in the diagnosis of lung disease.

[0147] In the example of Figure 8 the separating function is represented by a dashed line. In the examples of Figure 8 the function is a linear function forthat combination of features. In other examples, the functions may not be linear, they may be represented by non-linear functions. They may be other shapes. In some cases the function may be curved. In some cases the function may not be continuous, for example it may represent an area of the two-dimensional graph for the combination - for example a circle. In such examples, patients falling within the circle may belonging to a different group to those outside the circle. As described, the simplest method of combining features is a linear or a piecewise-linear approach, where the combined value is the weighted sum of each feature value multiplied by a coefficient which represents the contribution of the feature value to the combined value. A linear or piecewise-linear approach creates separating functions with straight dashed lines. The features can also be combined in a quadratic fashion. This approach assumes that the data within each class follows a multivariate normal distribution with different covariance matrices for each class. This approach creates a separating function that is curved. There are also non-parametric approaches for combining these features, such as using a decision tree, or ensembles of decision trees which create non-linear separating functions.

[0148] The examples of Figures 7 and 8 represent one-dimensional functions for a two-dimensional data set. In other examples, more than two functions may be combined to produce a combination-function having two 1 D functions. A combination-function can have multiple functions (e.g. 3, 4, 5 etc.). It will also be understood that for 3D data (e.g. where a compositefeature has 3 features) the function will be 2D, and that for 4D data the function will be 3D, and so on.

[0149] The examples of Figures 7 and 8 represent combinations of two features. In other examples, three dimensional combinations may be made. Examples of three-dimensional combinations include:

[0150] 1. OVI Skewness: This describes the asymmetry in the distribution of the oscillation / volatility of lung motion through time;

[0151] OVI Kurtosis: this describes the “tailedness” of the distribution of the oscillation / volatility of lung motion through time; and

[0152] Sum of high region of flow (t=2-5):

[0153] This describes the sum of proportion of the lung with high flow over time.

[0154] 2. OVI Skewness:

[0155] This describes the asymmetry in the distribution of the oscillation / volatility of lung motion through time;

[0156] OVI Kurtosis:

[0157] This describes the “tailedness” of the distribution of the oscillation / volatility of lung motion through time; and;

[0158] Sum of mean flow (t=1 -5): This describes the sum of mean air flow over time.

[0159] 3. OVI Skewness: This describes the asymmetry in the distribution of the oscillation / volatility of lung motion through time;

[0160] OVI Kurtosis: This describes the “tailedness” of the distribution of the oscillation / volatility of lung motion through time; and

[0161] Sum of mean flow (t=2-5): This describes the sum of mean air flow over time.

[0162] Figure 11 shows a graphical representation of a three-dimensional combination of features. The graphical representation of Figure 11 shows composite features derived from images taken on two groups of patients, a first control group and a second group having a lung condition (for example COPD). The composite feature is a combination of three features derived from patient image data. In Figure 11 , the axes represent the constituent features. In the example of Figure 11 , 3-dimensional separation function 11 10, in this case a plane (as a line can’t separate 3D data), separates those patients having the lung condition COPD (represented with pluses (+)) from those patients in the control group (represented by circles (o)). In the plot the with the x-axis showing a combined result of Feature B and Feature D, and the y-axis showing a projected view of Feature C (this is required to turn the separation function 1110, which in this instance is a plane, into a line on the plot). As can be seen from separation function 1 110, having a composite feature with three features can achieve greater separation in the data. An example of three features derived from patient image data is: ventilation distribution skewness (VDS); oscillation ventilation index (OVIhi) region; sum of flow heterogeneity (£FH).

[0163] For those combinations of features for which a separating function can be defined which distinguishes those patients having a lung condition from those without the lung condition, the relevant composite feature for a patient derived from lung image data can be compared with the separating function to predict whether a patient has the condition or not. If the composite feature falls on one side of the separating function, the system may predict that the patient has the lung condition, and if the composite feature falls on the other side of the separating function, the system may predict that the patient does not have the lung condition.

[0164] The particular features in the combinations provide information about the physical factors caused by the disease. For example, if one of the features is elasticity then this is an indication that the disease affects the elasticity of the lungs. By determining which combinations of features can be used to detect a particular lung condition, the features provide information about functional factors causing the lung condition. The comparison of the composite function with the separating function may generate a biomarker. The biomarker is a representation of the likelihood of that patient having the condition, and can therefore be used to predict whether the patient has the condition. For example, the value of the combined feature can optionally be passed through a non-linear function, such as a sigmoid function, to transform the combined value to have a range between 0 and 1 . In such an example 0 could represent a 0% likelihood of having a lung condition, and 1 could represent a 100% likelihood of having a lung condition. It will be understood that a biomarker is a measurable indicator of a biological state, such as the presence or severity of a disease. Biomarkers are used in medicine to identify and track various aspects of health and disease, including diagnosis and response to treatment.

[0165] Referring now to Figure 10, Figure 10a shows the same combined feature set as was shown in Figure 7c, and a separation function 1010 for differentiating between patients with Condition A or without condition A. Three patients 1011 , 1012, 1013 are shown, each of whom their composite feature falls on the Condition A side of the separation function 1010 (sometimes referred to as a decision boundary). Referring now to Figure 10b, in one embodiment, the biomarker for a patient may be the difference between the value of the composite function and the separating function. In graphical terms this may be the distance of the composite feature from the separating function (shown as 10121 for 1012 and 1011 1 for 101 1). The greater the distance the value of the composite feature is from the separating function, the higher the probability that the patient has that condition and so falls into the relevant group. For example, a patient having a composite feature very close to the separating function (e.g. 1013) may have a lower probability of actually having that condition than a patient having a composite function further from the separating function (e.g. patients 1011 , 1012, with patient 1012 being the furthest from the separation function, and therefore having the highest likelihood of having Condition A). In other words, although all patients 1011 , 1012, 1013 fall on the side of the separating function associated with having Condition A, the probability of Patient 1012 having the Condition A may be much higher than the probability of Patient 1013 having the lung condition. In this example, the value of the biomarker for Patient 1012 is different from the value of the biomarker of Patient 1013. The biomarker may be a representation of the distance of the value of the composite feature from the separating function.

[0166] Referring to Figure 10c, for some combination of features, the biomarker may be graded depending on the distance from the separating function. The system may include a predefined distance from the separating function. Composite features being positioned at a distance above the predefined distance may be provided at the same probability. For example, in Figure 10c there are three grading lines on each side separation function 1010, namely 10101 10102, 10103. The example of Figure 10c includes three grading lines 10101 10102 10103. 10101 being a first grading line and closest to separating function 10101 , grading line 10102 being further from separating function 1010, third grading line 10103 being still further from separating function 1010. These grading lines can be visual indicators, but they can also, if desired, define zones between the grading lines in which any values for the combined feature lying within the zone are given the same value for the biomarker. Alternatively, only the last grading line 10103 creates a grading zone, being a maximal grading zone, in which everything above that last grading line receives the maximum (or minimum if on the other side of the separating function 1010) value for the biomarker. In such an example Patient 101 1 and Patient 1013 would receive their raw value for the biomarker, whereas patients 1012 and 1014 would receive the maximum value for the biomarker. It will be understood that the grading lines may be equally or unequally spaced.

[0167] Biomarkers may be represented numerically. Biomarkers may be weighted differently for different lung diseases and for different combinations of features. For example, separating functions for some lung diseases may define an absolute distinction between groups. So, a composite function falling on one side of the separating function may indicate a very high probability of having a lung condition regardless of the distance from the separating function. Other separating functions for other lung conditions may define a less certain distinction between groups. In such cases, probability of a patient having a condition may only be significantly high above a predefined distance from the separating function.

[0168] Figures 9a and 9b show violin plots for the two patient groups represented in Figures 8a and 8b, respectively. In Figures 9a and 9b the biomarker is on the Y axis. In all cases a high score (close to 1) or low score (close to 0) indicate the strongest probability of the patient of being in one group (having the condition, 1) or the other (not having the condition, 0).

[0169] Figure 9a is a violin plot showing the biomarkers for the patients in Figure 8a using the composite feature of Feature B and Feature C. In Figure 9a the patients of the two groups are compared by showing the patients having Condition A in the left-hand side violin plot and patients in the control group in the right-hand side violin plot. Biomarkers are calculated using the separating function and the combination feature. This is achieved by passing the individual patient value of the combined feature through a non-linear function (e.g. a sigmoid function) to transform the combined value to have a range between 0 and 1 . In the example of Figure 9a the left side violin plot represents the Condition A patients. The number of the Condition A patients have a biomarker of greater than 0.8, and the large majority of Condition A patients have a biomarker of greater than 0.6. One patient has a 4D score of around 0.3. The right-side violin represents the control patients. The large majority of the control patients have a biomarker score of below 0.2, with almost all having a biomarker score of below 0.4. One patient has a biomarker of around 0.6. The good separation of the data in the violin plot demonstrates that this combination of features performs well, and that the Biomarker 1 could be used to detect the likelihood of having Condition A. The biomarker could be used simply as a score to inform a doctor on the likelihood, or it could be used in combination with an absolute threshold to make a definitive call on whether the patient does or does not have Condition A. With this combination of features there are likely to be some false positives and some false negatives, however this is expected in most medical testing.

[0170] In the example of Figure 9b the left side violin plot represents the Condition B patients and the right-side violin plot represents control patients. Figure 9b shows considerable overlap between the value of Biomarker 2 of patients in the two groups. This is an indication that the composite feature of Feature X and Feature Y is not a reliable indicator to diagnose Condition B.

[0171] As described above, Figures 9a and 9b show the spread of biomarkers for different groups of patients calculated from specific composite features for different lung conditions. The more distinct the separation in biomarkers between the groups, the greater the reliability of the separating function in detecting whether a patient has a particular lung condition or not, and so the more useful the composite feature is in diagnosing the lung condition. For example, in Figure 9a there is a clear difference between the biomarkers of the patients with Condition A and the control group, showing that the combination of features of Feature B and Feature C can provide a strong indication of whether a patient has a Condition A or not.

[0172] For example, to detect whether a patient has Condition A or not, patient image data can be used to generate a dataset capable of generating Feature B and Feature C, and then the value for Feature B and the value for Feature C can be calculated, and then those values combined to create a combined feature value. At this point the combined feature value can be used to determine whether the patient has Condition A or not (e.g. by plotting or comparing the values to the known separation function), or the combined feature value can be transformed into a biomarker, thereby providing a single number that represents either the likelihood of having the disease, or in combination with a single threshold provides an indication of whether the patient has Condition A. Embodiments derive composite features from patient image data and use the composite features to determine the presence of lung conditions.

[0173] An example implementation of the system for detecting lung disease is now described with reference to Figures 12 to 15.

[0174] As described above, lung image data 1310 for a patient is acquired typically using one of the techniques described above using imaging apparatus 1470, including X-Ray CT scans, paired X-ray CT scans, nuclear medicine imaging, X-ray fluorographs.

[0175] System 1400 is configured to perform the methods described to aid in the detection of lung disease. The components of system 1400 may be co-located or may be distributed and form part of a distributed computing system. In the case of a distributed computing system, components may be connected across communication networks, for example a mobile communication network. The distributed computing system may be implemented (at least in part) by a cloud-based computing environment, such as provided by Amazon Web Services (AWS), Microsoft Azure, Google Cloud Platform, and Oracle Cloud, in which the system 1400 is implemented as one or more virtual servers.

[0176] System 1400 may be controlled by user device 1490 via network 1480. For example, a clinician may access system 1400 via user device 1490 to diagnose whether or not a patient has a particular lung condition. The clinician may have access to image data for a patient and may use system 1400 to analyse that image data to determine whether or not the patient has a particular lung condition. The user device may be a computing device with a keyboard, touchscreen, touch pad, microphone or other input device.

[0177] System 1400 may include one or more network interfaces 1410 that facilitate communication between the system 1400 and one or more other apparatuses using any suitable communications standard via network 1480. For example, network interface 1410 may enable the receipt of image data from imaging apparatus 1470, where the image data represent images captured by the imaging apparatus 1470. The interface 1410 may also enable the receipt of further information relating to the patient and / or the image datasets. Interface 1410 may enable the receipt of datasets derived from the image data.

[0178] The interface 1410 may be a LAN interface that implements protocols and / or algorithms that comply with various communication standards of the Institute of Electrical and Electronics Engineers (IEEE), such as IEEE 802.11 , while a cellular network interface implement protocols and / or algorithms that comply with various communication standards of the Third Generation Partnership Project (3GPP) and 3GPP2, such as 3G and 4G (Long Term Evolution), and of the Next Generation Mobile Networks (NGMN) Alliance, such as 5G. The image data may be acquired at the time of analysis, or acquired in an imaging session taken at an earlier time.

[0179] Processor 1420, implemented in hardware may be a general-purpose processor. A general- purpose processor may be a microprocessor, but in the alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor 1420 may include, without limitation, a central processing unit (CPU), a digital signal processor (DSP), a reduced instruction set computer (RISC) processor, a complex instruction set computer (CISC) processor, a microprocessor, a microcontroller, a field programmable gate array (FPGA), a System-on-a-Chip (SOC), or other programmable logic, discrete gate or transistor logic, discrete hardware components, or any combination thereof, or any other suitable component designed to perform the functions described herein. Processor 1420 may also include one or more application- specific integrated circuits (ASICs) or application -specific standard products (ASSPs) for handling specific data processing functions ortasks. Processor 1420 may also be implemented as a combination of computing components, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP, or any other such configuration.

[0180] Software or firmware implementations of processor 1420 may include computer executable or machine-executable instructions written in any suitable programming language to perform the various functions described herein. Software shall be construed broadly to mean instructions, instruction sets, code, code segments, program code, programs, subprograms, software modules, applications, software applications, software packages, routines, subroutines, objects, executables, threads of execution, procedures, functions, etc., whether referred to as software, firmware, middleware, microcode, hardware description language, or otherwise. The software may reside on a computer-readable medium. A computer-readable medium may include, by way of example, a smart card, a flash memory device (e.g., card, stick, key drive), random access memory (RAM), read only memory (ROM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), a general register, or any other suitable non-transitory medium for storing software.

[0181] Memory 1430 may include, but is not limited to, random access memory (RAM), flash RAM, magnetic media storage, optical media storage, and so forth. The memory 1430 may include volatile memory configured to store information when supplied with power and / or non-volatile memory configured to store information even when not supplied with power. The memory 1430 may store various program modules, application programs, and so forth that may include computer-executable instructions that upon execution by the processor 1420 may cause various operations to be performed. The memory 1430 may further store a variety of data manipulated and / or generated during execution of computer-executable instructions by the processor 1420.

[0182] Memory 1430 may store various program modules, application programs, and so forth that may include computer-executable instructions that upon execution by the processor 1420 may cause various operations to be performed.

[0183] Memory 1430 includes storage modules, and stores applications for aiding in the detection of lung disease. Each of the modules includes functions in the form of logic and rules that respectively support and enable the various functions described herein with reference to the Figures.

[0184] User interface 1440 facilitates user interaction with apparatus 1400. User interface includes a user input module to receive input from the user and provide output to a user. The user interface 1440 may communicate with a user device 1490 via network 1480.

[0185] At 1210 patient image data is acquired by the lung condition analysis system 1400. The image data is acquired at the network interface 1410.

[0186] At 1220 1 1320 a dataset is derived from the image data using processor 1420. The dataset may be derived using any of the methods described above. The type of dataset may be selected by the user via user device 1490. The system 1400 receives a request to receive a particular type of dataset. The dataset may be structural dataset identifying the structure of the lung. The dataset may be a dataset describing lung function, for example a regional perfusion dataset describing the blood flow in the lung, or a regional ventilation dataset describing the air in the lung.

[0187] At 1230 1 1330 processor 1420 derives two or more features of the dataset. A feature is a 1- dimensional representation of the data points of the dataset. The features may be selected by the user via user device 1490. As described above, particular features and composite features may be useful in detecting particular types of lung disease. The user may select specific features depending on which lung disease is to be diagnosed. At 1240 / 1340 processor 1430 combines the two or more features of the dataset to produce a composite feature. The combination may be a predefined combination of features known to indicate the presence of a lung disease. Predefined combination of features may be stored in memory 1430. The predefined combinations are associated with particular lung diseases. For example, a particular lung condition may be known to be detectable from a specific combination of features.

[0188] At 1250 the system uses the composite feature to detect the presence of lung disease. As described above, this step may be performed by retrieving a separation function associated with the composite feature and associated with a particular lung disease. By comparing the composite feature value for the patient dataset the system can determine the likelihood of the patient having the particular lung disease.

[0189] The system may generate a biomarker from the comparison of the composite feature and the separation feature. The biomarker may be used to predict the likelihood of the patient having the particular lung disease.

[0190] Memory 1430 may store data defining at least one lung condition and composite functions associated with that lung condition that enable the presence of that lung condition to be detected. Table 1 provides a non-limiting example of data that may be stored in memory 1430.

[0191] In the example of Table 1 , various lung conditions are identified which can be detected using composite features derived from image data. It will be appreciated that not all lung conditions may be detectable using this composite feature technique and so not all lung conditions may be listed in the table. Table 1 identifies lung conditions that can be detected is identified using composite features, i.e. Lung Condition A, B, C etc. The Lung Conditions are listed alphabetically in the example of Table 1 for illustrative purposes only. In practice, Lung Condition A may refer to COPD, Lung Condition B may refer to Asthma, etc.

[0192] For each condition, the features required for the relevant composite feature are stored. Table 1 identifies that the presence of Lung Condition A can be detected using a composite feature, being a combination of Features 1 , 3 and 8. Forthe purposes of illustration, in Table 1 different features are labelled numerically. In practice, Feature 1 may refer to OVI Kurtosis, Feature 7 may refer to Maximum flow skewness (t=2-4), etc.

[0193] Additional information may be included in the memory. For example, memory 1430 may identify the image data required in order to derive the required feature. For example: Feature 15 may be derivable from a paired CT scan; Feature 21 may be derivable from a single CT scan at max inhalation; Feature 9 may be derivable from a nuclear VQ scan. This information may be important for the clinician or other analysing party so they are made aware of what scans are required in order to detect that Lung Condition. For example, if a nuclear CT image data is not available for a patient then the clinician would be unable to use the system to detect the presence of Lung Condition E since Feature 9 could not be derived because the required scan (nuclear VQ scan) is not available.

[0194] Table 1 includes the separating function for the specific composite function and lung disease. The separating function can be compared with the composite feature.

[0195] Table 1 also includes information about the biomarker. The biomarker can be compared with the comparison of the separating function and composite feature to determine the likelihood of the patient having the lung condition. As described above, the biomarker for a patient may be the difference between the value of the composite function and the separating function. The biomarker may be calculated by passing the value of the combined feature through a nonlinear function, such as a sigmoid function, to transform the combined value to have a range between 0 and 1 .

[0196] Referring now to Figure 15, a user may select a lung condition for diagnosis. System 1400 receives a request to diagnose a lung condition at 1510 at user interface 1440. At 1520 system 1400 determines which features are required to produce a composite feature associated with the requested lung condition. The system determines what image data is required to derive the required features. At 1530 the system retrieves the required image data. System 1400 determines what data sets are required to be derived from the image data to generate the required features. At 1540 system 1400 derives the required data sets from the image data. At 1550 system 1400 derives the two or more required features from the datasets and calculates the value of the two or more features. System 1400 combines the value of the features to produce a composite feature value at 1560.

[0197] System 1400 retrieves the separating function associated with the selected lung disease and composition feature. At 1570 system 1400 compares the composite feature value with the separating function. This result of the comparison may be presented to the user via user interface 1440, for example it may be presented graphically in a graph similarto that of Figures 7c, for example with features on different axes and the separation function presented. In other examples, the result of the comparison may be provided in other graphical or non-graphical forms to the user via user interface 1440. Alternatively, the system can generate a report, for example in PDF format, with the information.

[0198] At 1580 system 1400 may calculate a biomarker value from the composite feature value and the separation function. The biomarker may be used to determine the likelihood of the presence of the lung condition, for example it may provide the probability of the patient having the lung condition.

[0199] The system 1400 may work under the control of user device 1490 but may also perform the steps automatically. The system work as a combination of automatic steps and steps controlled by used device 1490.

[0200] Embodiments can be used to detect lung conditions. The lung conditions may be a lung disease. The lung disease may be a disease caused by a patient breathing toxic hazards, to which certain parts of the population may be exposed, for example coal miners or firefighters. The lung condition may be caused by particular toxic hazards, for example DRRD.

[0201] Lung conditions may also include pneumothorax.

[0202] It is to be understood that, if any prior art publication is referred to herein, such reference does not constitute an admission that the publication forms a part of the common general knowledge in the art, in Australia or any other country. In the claims which follow and in the preceding description of the invention, except where the context requires otherwise due to express language or necessary implication, the word "comprise" or variations such as "comprises" or "comprising" is used in an inclusive sense, namely, to specify the presence of the stated features but not to preclude the presence or addition of further features in various embodiments of the invention.

[0203] It is to be understood that the aforegoing description refers merely to preferred embodiments of invention, and that variations and modifications will be possible thereto without departing from the spirit and scope of the invention, the ambit of which is to be determined from the following claims.

Claims

Claims:1 . A method for use in diagnosing a respiratory health condition of a patient comprising the steps of: providing image data of lungs of a patient; deriving two or more features from the image data; combining the two or more features to produce a composite feature; using the composite feature to diagnose the presence of a respiratory health condition in the lungs of the patient.

2. A method according to claim 1 further comprising the steps of selecting a respiratory health condition for diagnosis, identifying a composite feature associated with the selected respiratory health condition, the composite feature comprising at least two features, wherein the derived two or more features at the at least two features comprised of the composite feature.

3. A method according to claim 1 or 2 further comprising providing a separation function associated with the composite feature, the separation function, wherein the step of diagnosing the presence of the respiratory health condition is performed by comparing the composite feature with the separation function.

4. A method according to claim 1 , 2 or 3 comprising a further step of calculating a biomarker, the biomarker being the comparison of the composite feature with the separation function.

5. A method according to claim 4 wherein the biomarker being an indication of the likelihood of the patient having the respiratory health condition.

6. A method according to any of claims 1 to 5 comprising the further step of deriving a dataset from the image data, wherein the two or more features are derived from the dataset.

7. A method according to claim 6 comprising the step of extracting the value of the two or more features and the step of combining the features is performed by combining the values of the features to produce a composite feature value, and use the compositefeature value to diagnose the presence of a respiratory health condition in the lungs of the patient.

8. A method according to any one preceding claim where a feature is one-dimensional representation of the image data.

9. A method according to any one preceding claim wherein the features are at least one of: functional features, defining a function of the lungs; and / or, structural features, defining structure of the lungs.

10. A method according to any one preceding claim wherein the features include temporal and spatial information.

11. A method according to claim 6 wherein the dataset is at least one of: a structural dataset; ventilation dataset; oscillation dataset; flow dataset; perfusion dataset.

12. A method according to any preceding claim wherein the respiratory health condition comprises: collateral ventilation disease, small airways disease.

13. A method according to any preceding claim wherein the features provide information about functional factors causing the lung condition14. A software application stored on a non-transitory medium adapted to perform the method of any one preceding claim.

15. A system configured for use in diagnosing a respiratory health condition of a patient comprising: processor configured to: provide image data of lungs of a patient; derive two or more features from the image data; combine the two or more features to produce a composite feature; use the composite feature to diagnose the presence of a respiratory health condition in the lungs of the patient.

16. A method for use in diagnosing a respiratory health condition of a patient comprising the steps of:acquiring image data of the lungs from at least two patients, wherein at least one of the at least two patients has a first lung condition and at least one other of the at least two patients does not have the lung condition; deriving two or more features from the image data for the at least two patients; combining the two or more features for each patient to produce a composite feature for each patient; generating a separating function, the separating function being a mathematical function of the derived two or more features which separates those patients of the at least two patients having the first lung condition from those patients of the at least two patients not having the condition; the separating function for use in diagnosing a respiratory health condition.

17. A method according to claim 3, wherein the separating function is generated according to the method of claim 16.

18. A system for use in diagnosing a respiratory health condition of a patient comprising: processor configured to: acquire image data of the lungs from at least two patients, wherein at least one of the at least two patients has a first lung condition and at least one other of the at least two patients does not have the lung condition; derive two or more features from the image data for the at least two patients; combine the two or more features for each patient to produce a composite feature for each patient; generate a separating function, the separating function being a mathematical function of the derived two or more features which separates those patients of the at least two patients having the first lung condition from those patients of the at least two patients not having the condition; the separating function for use in diagnosing a respiratory health condition.

Images