Gamma feature localisation

Abstract

A computer-implemented method for localising gamma features in a gamma image obtained (210) by detecting gamma radiation is provided. The method comprises: passing (220) the gamma image through a mach

Description

TECHNICAL FIELD The present invention relates generally to a computer-implemented method of localising gamma features in images obtained by detecting gamma radiation, methods for training machine learning models for localising gamma features in gamma images and to corresponding systems. BACKGROUND Gamma imaging has many uses in medical imaging but also during clean-up of radioactively contaminated sites and other scenarios. It may be advantageous to provide handheld and lightweight gamma imaging cameras for handheld operation by a user. When compared to larger static gamma imaging systems, portable systems detect fewer incident gamma particles due to having lower sensitivity and shorter imaging times. This leads to sparse data (on the order of fewer than 100 counts per frame), substantial changes from frame to frame due to the stochastic nature of gamma radiation, and poor spatial resolution. Machine learning models may be used to detect features in high quality images with dense features, such as those obtained using X-rays, PET scans, and MRIs. In gamma images, typically produced by large static gamma imaging systems, regions of interest may be manually identified corresponding with radioactive sources. Localising gamma features in data obtained from handheld portable gamma cameras remains challenging. It is desirable to be able to detect gamma features automatically and objectively in low count gamma images, such as those produced by handheld portable gamma cameras. STATEMENTS OF INVENTION According to a first aspect of the invention, there is provided a computer-implemented method for localising gamma features in a gamma image obtained by detecting gamma radiation. The method comprises passing the gamma image through a machine learning model. The machine learning model is configured to identify bounding boxes which indicate possible locations of gamma features. When the machine learning model has identified at least one bounding box, the method further comprises: deriving at least one metric for each of the at least one bounding box based on the bounding box and the gamma image; and outputting the bounding boxes that satisfy a predetermined constraint of the at least one metric. The term “gamma feature” may mean a potential source of gamma radiation. The term “bounding box” may mean a rectangular set of pixels in the gamma image which indicate that a possible location of a gamma feature is located within that set. In general, a bounding box need not be a rectangular set and may be in general any convex set of pixels surrounding a possible location of a gamma feature. The gamma image may be obtained by a gamma camera comprising a scintillator and a photodetector. The scintillator may be configured to produce light in response to incident gamma radiation. The photodetector may be arranged to detect the light produced by the scintillator in response to the incident gamma radiation. In some alternative embodiments, the gamma camera may be a solid-state gamma detector or a direct gamma detector. The gamma camera may weigh less than 10kg, or 5kg, or 2kg. In some embodiments, the gamma camera is sufficiently light so that an operator may lift it without external support. The gamma camera may have a field of view at least as wide as 60°. The gamma camera may be a hybrid gamma camera, operable to produce a gamma image obtained by detecting gamma radiation and an optical image (e.g. obtained by imaging light in the visible spectrum). Alternatively, the optical image may be obtained by imaging light in the infrared spectrum, or another imaging modality. The gamma image may superimpose the detected gamma radiation onto the optical image. The at least one metric may describe feature size, feature density, and / or image quality. Feature size may describe the size of the bounding box. The feature size may be a relative feature area size or an absolute feature area size. Relative feature area size may be the proportion of the number of pixels in the bounding box to the number of pixels in the gamma image. The absolute feature area size may be the number of pixels in the bounding box. Feature density may be the average number of incident gamma radiation particles per pixel within the bounding box. The image quality may be a contrast to noise ratio. The contrast to noise ratio may be the average number of incident gamma radiation particles within the bounding box divided by the standard deviation of the background noise outside the bounding box. Other metrics, such as signal to noise ratio, peak signal to noise ratio, mean squared error, and structural similarity, are equally envisaged. The machine learning model may be a single shot object detection model. The machine learning model may be a YOLO network, such as a YOLOv7-X network. According to a second aspect of the invention, there is provided a computer-implemented method for training a machine learning model to localise gamma features in a gamma image. The method comprises: a) identifying at least one bounding box which indicates possible locations of gamma features by providing the gamma image to a machine learning model; b) computing a loss by comparing the at least one bounding box with an at least one known bounding box; c) backpropagating the loss through the machine learning model to determine updated model parameters of the machine learning model; d) repeating a) to c) using a first plurality of gamma images until a predetermined criterion is met; e) evaluating the machine learning model by providing a second plurality of gamma images to the machine learning model, and, for each gamma image of the second plurality of gamma images, computing and storing at least one metric; f) determining a constraint based on the stored at least one metric. Gamma images, bounding boxes, machine learning model, and at least one metric may be as described according to the first aspect of the invention. The loss may be an objectness loss, box prediction loss, and / or a classification loss or a combination thereof. Backpropagating the loss through the machine learning model may comprise using stochastic gradient descent to optimise the machine learning parameters based on the loss. The predetermined criterion in d) may be repeating a) to c) a predetermined number of times; repeating a) to c) until the loss is below a predetermined threshold; or any other appropriate early stopping criterion. Evaluating the machine learning model may further comprise determining if a bounding box has been found and determining if the bounding box is within a predetermined distance to a known bounding box. Having a bounding box within a predetermined distance of a known bounding box may mean having an intersection over union (loU), positional accuracy, localisation accuracy, classification accuracy, and / or localisation precision over a predefined value. Having a bounding box within a predetermined distance of a known bounding box may alternatively mean having the centre of the bounding box to be within a predetermined distance of a centre of the known bounding box. Said predetermined distance may depend on the distance of the gamma source to the gamma camera. Determining a constraint may comprise modelling at least one relationship between the at least one metric according to the first aspect of the invention and at least one of loU, positional accuracy, localisation accuracy, and / or localisation precision; using the at least one modelled relationship to determine constraints that result in the at least one of loU, positional accuracy, localisation accuracy, and / or localisation precision, being over a threshold. According to a third aspect of the invention, there is provided a system comprising one or more processors. The one or more processors are configured to perform the steps of: passing a gamma image obtained by detecting gamma radiation through a machine learning model, wherein the machine learning model is configured to identify bounding boxes which indicate possible location of gamma features; wherein when the machine learning model has identified at least one bounding box: deriving at least one metric for each of the at least one bounding box based on the bounding box and the gamma image; outputting the bounding boxes that satisfy a predefined constraint of the at least one metric. The system of the third aspect may be configured to perform any of the method steps described with reference to the first aspect, including optional features thereof. According to a fourth aspect of the invention, there is provided a gamma feature detection system comprising a gamma camera and the system according to the third aspect of the invention. The gamma camera is configured to capture a gamma image obtained by detecting gamma radiation and send the gamma image to the system according to the third aspect of the invention. The system according to the third aspect of the invention is further configured to receive the gamma image. The gamma camera may comprise a scintillator and a photodetector. The scintillator may be configured to produce light in response to incident gamma radiation. The photodetector may be arranged to detect the light produced by the scintillator in response to the incident gamma radiation. In some alternative embodiments, the gamma camera may be a solid-state gamma detector or a direct gamma detector. The gamma camera may weigh less than 10kg, or 5kg or 2kg. In some embodiments, the gamma camera is sufficiently light so that an operator may lift it without external support. The gamma camera may have a field of view at least as wide as 60°. There may be provided a computer program, which when run on a computer, causes the computer to configure any apparatus, including a circuit, controller, sensor, filter, or device disclosed herein or perform any method disclosed herein. The computer program may be a software implementation, and the computer may be considered as any appropriate hardware, including a digital signal processor, a microcontroller, and an implementation in read only memory (ROM), erasable programmable read only memory (EPROM) or electronically erasable programmable read only memory (EEPROM), as non-limiting examples. The software implementation may be an assembly program. The computer program may be provided on a non-transitory computer readable medium, which may be a physical computer readable medium, such as a disc or a memory device, or may be embodied as a transient signal. Such a transient signal may be a network download, including an internet download. These and other aspects of the invention will be apparent from, and elucidated with reference to, the embodiments described hereinafter. BRIEF DESCRIPTION OF FIGURES Embodiments will be described, by way of example only, with reference to the drawings, in which: Figure 1 is an example of two gamma images and a hybrid gamma camera; Figure 2 is a flow diagram for localising gamma features in a gamma image; Figure 3 is a flow diagram for training and evaluating a machine learning model to localise gamma features in a gamma image; Figure 4 is a plurality of gamma images with predicted and ground truth bounding boxes for localised gamma features; Figure 5 is a plurality of gamma images with predicted and ground truth bounding boxes for localised gamma features; and Figure 6 is a plurality of gamma images obtained from patients with predicted bounding boxes for localised gamma features. It should be noted that the Figures are diagrammatic and not drawn to scale. Relative dimensions and proportions of parts of these Figures have been shown exaggerated or reduced in size, for the sake of clarity and convenience in the drawings. The same reference signs are generally used to refer to corresponding or similar features in modified and different embodiments. DETAILED DESCRIPTION The problem of object detection in gamma images taken using relatively small gamma cameras, such as the hybrid gamma cameras (hereinafter HGC) depicted in Figure 1 is described with reference to Figure 1. Although a HGC is depicted, it is not essential that gamma cameras in embodiments are provided with an optical (e.g. visible light) camera in addition to a gamma detector. The teaching of the present application may equally be applied to any relatively small gamma camera, such as the sort of devices that are referred to in the literature as small field of view (SFOV) gamma cameras or a handheld gamma camera. Equally, the application may be applied to large field of view cameras if the scenario of use results in an equivalent sparsity of counts in the image. Figure 1 shows a clinical gamma image 10 of a patient’s thyroid obtained a standard clinical gamma camera. Figure 1 further shows a sparse gamma image 12 of the same area obtained using an HGC 100. In the clinical gamma image 10, the left and right lobes of the thyroid 101, 103 are clearly visible. In the sparse gamma image 12, each white pixel represents detected gamma radiation (e.g., a photon of gamma radiation). There is no clear structure to the detected radiation and so it is not possible to precisely localise the thyroid. The HGC 100 is provided with an optical camera 102, an angled mirror 104, a pinhole collimator 106, a scintillator 108, and a photodetector 110. The optical camera 102 and angled mirror 104 are configured to capture optical images and / or video so that the gamma features can be superimposed onto the optical images and / or video to allow for user clarity. The pinhole collimator 106 maps the incoming gamma radiation from different directs onto different locations of the scintillator 108. In turn, the scintillator 108 produces light in response to incident gamma radiation. The light produced by the scintillator 108 is detected by the photodetector 110 which then registers the light as the sparse gamma image 12. With reference to Figure 2, there is shown a flow diagram for localising gamma features in a gamma image. At 210, a gamma image is obtained by detecting gamma radiation using the HGC 100 (in principle the techniques here are applicable to any gamma image but are especially relevant to sparse gamma images of the sort that are produced by HGCs). The scintillator is configured to produce light in response to incident gamma radiation, and the photodetector 110 is arranged to detect the light produced by the scintillator 108. The HGC 100 weighs approximately 1.5kg and so can be operated and hand-held by a user without the need of any external support. Advantageously, using a portable hand-held camera is cheaper than requiring large static gamma detectors in a clinical environment and is also more convenient during clean-up of radioactively contaminated sites. At 220, the obtained gamma image is passed through a single shot object detection model to identify at least one bounding box. In the present embodiment, the single shot object detection model is a Y0L0v7-X model. Advantageously, YOLO and other single shot object detection models can process images at higher frame rates when compared to other object detection models when run on the same hardware. In the present embodiment, the object detection model is run on a desktop computer or a rugged tablet. Optionally, the desktop computer may have GPU accelerators. Further optionally, the rugged tablet may use cloud servers to run the detection model. This allows for relatively short delays when localising gamma features. This provides a more convenient method that can provide a user with locations of gamma sources in nearly real time during, for example, a clean-up of radioactively contaminated sites. The skilled person will understand that other single shot object detection models are equally envisaged. The model is trained to produce bounding boxes that correspond to possible locations of gamma features. In the present embodiment, gamma features are sources of gamma radiation, such as a radioactive object or a patient’s thyroid after uptake by a radiotracer. In some embodiments, the model is trained to over-identify bounding boxes in order to minimise the number of false negatives. False positives are further triaged at 230 and 240. At 230, at least one metric for each of the bounding boxes identified at 220 are derived. In some embodiments, the relative feature area size (Ar), absolute feature area size (Apb), feature density (p), and contrast to noise ratio (CNR) are derived for each bounding box. Ar is the proportion of the bounding box to the entire gamma image size, i.e., the number of pixels in the bounding box divided by the number of pixels in the gamma image. Apb is the number of pixels in the bounding box. p is the average number of counts per pixel within the bounding box. That is, the sum of all the pixel values divided by the number of pixels. The CNR is the average number of incident gamma radiation particles within the bounding box divided by the standard deviation of the background noise outside the bounding box. Advantageously, these four metrics are relatively simple to compute which are more effective than more complicated metrics given the sparsity of detected gamma radiation. At 240, the metrics calculated for each bounding box are compared to a predefined constraint. The bounding boxes whose metrics satisfy the predefined constraints are then outputted. In some embodiments, the predefined constraints are a set of logically connected inequalities. For example, the predefined constraints could be Ar >0.05, or Apb >40000 with Ar <0.05, and p >0.05, and CNR >0.5. These example values for predefined constraints and are not intended to be limiting and other values may also be used. A method of determining suitable predefined constraints will be described below. Turning to Figure 3, there is shown a flow diagram for training a machine learning model to localise gamma features in a gamma image. Due to lack of available real life patient data, training, validation, and evaluation datasets are simulated using Monte Carlo physics simulations. Different types of gamma sources are simulated and the gamma image that would be detected by the HGC is calculated. This allows for more data as well as known ground truth bounding boxes of the gamma sources for each gamma image. At 310, an untrained machine learning model identifies at least one bounding box in one of the gamma images from the training dataset. As described above, the machine learning model is a single shot object detection model, preferably YOLOv7-X. At 320, a loss is computed based on the identified at least one bounding box and the ground truth bounding boxes. The loss may be one of objectness loss, box prediction loss, or classification loss, or a combination thereof. Any appropriate loss used in object detection models may be appropriate and can be selected on an ad-hoc basis during hyper-parameter optimisation. At 330, the loss computed in 320 is backpropagated through the machine learning model in order to determine updated parameters of the machine learning model in an attempt to minimise loss. In some embodiments, this is done using stochastic gradient descent. However, the skilled person will recognise that any suitable optimisation algorithms, such as ADAM, AdaGrad, or RMSProp, may equally be suitable. At 340, the machine learning model checks if a predetermined criterion is met, i.e., whether the machine learning model has reached a satisfactory level of training. In the present embodiment, the predetermined criterion is repeating 310, 320, and 330 a predetermined number of times (e.g., 100 times). Alternatively, the predetermined criterion may be repeating 310, 320, and 330 until the loss computed at 320 is below a predetermined threshold. The skilled person will understand that any suitable early stopping criterion may be appropriate. If the predetermined criterion is not met, then the method proceeds with 345, else the method proceeds with 360. If the predetermined criterion is not met, then at 345, the machine learning model repeats 310, 320, 330 with a different gamma image (or batch of gamma images) from the training dataset. If the predetermined criterion is met, then at 350 a plurality of gamma images from an evaluation dataset are passed through the object detection model. For each of the bounding boxes detected, at least one metric of the bounding box is computed and stored. The at least one metric includes all the metrics described in relation to 230 and also includes positional accuracy. Positional accuracy is defined as the distance between the centre of the ground truth bounding box and the identified bounding box. The positional accuracy is preferred over the intersection over union (loU) since positional accuracy only considers the position rather than the position and size which are both factors in the loU. By passing gamma images (or batches of gamma images) from the evaluation dataset through the machine learning model, a Localisation Accuracy (LA) is computed. LA is computed by taking the number of gamma images with predicted bounding boxes having a positional accuracy within a predefined margin and dividing it by the number of gamma images. The predefined margin may depend on the simulated geometry of Monte Carlo physics simulations, such as the simulated imaging distance. It has been realised that LA exhibits a dependency on Ar,Apb,p, andCNR. At 360, the relationship between LA and each of Ar,Apb,p, and CNR are individually modelled using linear bivariate splines. Using these linear bivariate splines, constraints on Ar,Apb,p, andCNR can be determined to ensure that LA is sufficiently high. These constraints then become the predefined constraints used in 240. Turning to Figures 4 and 5, there is shown test data using the computer-implemented method of localising gamma features according to the present disclosure. Gamma images 401, 403, 405, 501, and 503 show simulated gamma images with relatively few gamma counts (23, 13, 20, 14, and 22 counts, respectively) that were obtained as part of the test data for evaluating the performance of the model. This level of gamma counts is on the order of what an HGC measures when in use. Gamma images 402, 404, 406, 502, and 504 show simulated gamma images with the same respective gamma source but with much higher gamma counts. The respectively higher count subfigures are comparable to what could be obtained using larger static gamma imaging systems. Using the method according to the present disclosure, predicted bounding boxes 411, 413, 415, 511, 513 that predict the locality of the gamma source are shown. These predicted bounding boxes 411, 413, 415, 511, 513 can be compared to the ground truth bounding boxes 412, 414, 416, 512, 514, respectively. The predicted bounding boxes 411, 413, 415, 511, 513 correspond closely to the ground truth bounding boxes 412, 414, 416, 512, 514. This shows that the model works well at localising gamma features with as few as 13 gamma counts. Furthermore, turning specifically to gamma image 503, there can be seen outlier gamma counts 534 (circled white for clarity). These outlier gamma counts 534 were not included within the predicted bounding box 513 by the machine learning model. This shows that the machine learning model does more than mere naive implementation and is able to construct bounding boxes which are most likely to assimilate the ground truth bounding boxes. Turning to Figure 6, there is shown a plurality of gamma images 601, 603, 605, 607 acquired from patients using an HGC Similarly, the computer-implemented method described in Figure 2 was applied to identify predicted bounding boxes 602, 604, 606, 608, 610, 612, 614. Predicted bounding boxes 604, 608, 612, and 614 are provided with the respective default prediction confidence output by the YOLO network. The bounding boxes with higher feature densities indicate that it is more likely that the bounding box corresponds to a gamma feature. These gamma images 601, 603, 605, 607 and corresponding predicted bounding boxes 602, 604, 606, 608, 610, 612, 614 show that the computer-implemented method described in Figure 2 has practical uses in a medical environment. From reading the present disclosure, other variations and modifications will be apparent to the skilled person. Such variations and modifications may involve equivalent and other features which are already known in the art of machine learning, and which may be used instead of, or in addition to, features already described herein. Although the appended claims are directed to particular combinations of features, it should be understood that the scope of the disclosure of the present invention also includes any novel feature or any novel combination of features disclosed herein either explicitly or implicitly or any generalisation thereof, whether or not it relates to the same invention as presently claimed in any claim and whether or not it mitigates any or all of the same technical problems as does the present invention. Features which are described in the context of separate embodiments may also be provided in combination in a single embodiment. Conversely, various features which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable sub combination. The applicant hereby gives notice that new claims may be formulated to such features and / or combinations of such features during the prosecution of the present application or of any further application derived therefrom. For the sake of completeness, it is also stated that the term “comprising” does not exclude other elements or steps, the term “a” or “an” does not exclude a plurality, a single processor or other unit may fulfil the functions of several means recited in the claims and reference signs in the claims shall not be construed as limiting the scope of the claims.

Claims

1. A computer-implemented method for localising gamma features in a gamma image obtained by detecting gamma radiation, the method comprising:passing the gamma image through a machine learning model, wherein the machine learning model is configured to identify bounding boxes which indicate possible locations of gamma features;wherein when the machine learning model has identified at least one bounding box:deriving at least one metric for each of the at least one bounding box based on the bounding box and the gamma image;outputting the bounding boxes that satisfy a predefined constraint of the at least one metric.

2. The computer-implemented method of claim 1, wherein the gamma image is obtained by a gamma camera comprising a scintillator and a photodetector, the scintillator configured to produce light in response to incident gamma radiation, and the photodetector is arranged to detect the light produced by the scintillator in response to the incident gamma radiation.

3. The computer-implemented method of claim 2, wherein the gamma camera weighs less than 2kg.

4. The computer-implemented method of any preceding claim, wherein the gamma camera has a field of view at least as wide as 60°.

5. The computer-implemented method of any preceding claims, wherein the at least one metric describes a feature size, feature density, and / or image quality.

6. The computer-implemented method of any preceding claim, wherein the at least one metric is at least one of: relative feature area size, absolute feature area size, feature density, and / or contrast to noise ratio.

7. The computer-implemented method of any preceding claim, wherein the machine learning model is a single shot object detection model.

8. A computer-implemented method for training a machine learning model to localise gamma features in a gamma image, the method comprising:a) identifying at least one bounding box which indicates possible locations of gamma features by providing the gamma image to a machine learning model;b) computing a loss by comparing the at least one bounding box with an at least one known bounding box;c) backpropagating the loss through the machine learning model to determine updated model parameters of the machine learning model;d) repeating a) to c) using a first plurality of gamma images until a predetermined criterion is met;e) evaluating the machine learning model by providing a second plurality of gamma images to the machine learning model, and, for each gamma image of the second plurality of gamma images, computing and storing at least one metric;f) determining a constraint based on the stored at least one metric.

9. A system comprising one or more processors, configured to perform the steps of:passing a gamma image obtained by detecting gamma radiation through a machine learning model, wherein the machine learning model is configured to identify bounding boxes which indicate possible locations of gamma features;wherein when the machine learning model has identified at least one bounding box:deriving at least one metric for each of the at least one bounding box based on the bounding box and the gamma image;outputting the bounding boxes that satisfy a predefined constraint of the at least one metric.

10. A gamma feature detection system comprising a gamma camera and a system according to claim 9, wherein:the gamma camera is configured to capture a gamma image obtained by detecting gamma radiation and send the gamma image to the system according to claim 9; andthe system of claim 9 is further configured to receive the gamma image.

11. The gamma feature detection system of claim 10, wherein the gamma camera comprises a scintillator and a photodetector, the scintillator configured to produce light in response to incident gamma radiation, and the photodetector is arranged to detect the light produced by the scintillator in response to the incident gamma radiation.

12. The gamma feature detection system of any of claims 10 or 11, wherein the gamma camera weighs less than 2kg.

13. The computer-implemented method of any of claims 10 to 12, wherein the gamma camera has a field of view at least as wide as 60°.

14. A non-transient machine readable medium comprising instructions for configuring one or more processors to perform the steps of:passing the gamma image through a machine learning model, wherein the machine learning model is configured to identify bounding boxes which indicate possible locations of gamma features;wherein when the machine learning model has identified at least one bounding box:deriving at least one metric for each of the at least one bounding box based on the bounding box and the gamma image;outputting the bounding boxes that satisfy a predefined constraint of the at least one metric.

Images