DIRECT SKIN RADIATION DOSE MEASUREMENT SYSTEM WITH QUANTITATIVE OPTICAL
READ-OUT
Field of the invention
The invention relates to the field of electromagnetic as well as particle radiation dosimetry. More specifically it relates to systems and methods for non-invasive radiation dosimetry on external surfaces in general and, in particular, for in-vivo skin dosimetry upon exposure to UV or ionizing radiation, and to a device combined with optical read-out method used in said system.
Background of the invention
Exposure to different forms of electromagnetic and particle radiation is commonly experienced by the vast majority of the population. Relevant sources of exposure are UV radiation for the general population and medical radiation (X-rays, gamma rays, etc.) for patients and medical personnel. This has been documented in different studies as, for example, "Health risks of Ionizing Radiation: an Overview of Epidemiological Studies"; A report by the Community-Based Hazard Management Program, George Perkins Marsh Institute, Clark University, March 2006; and "X-rays, Gamma Rays and cancer risk'ç Medical Review of the American CancerSociety, 2013.
Ionizing radiation is commonly used in medical practice for diagnosis, interventional guidance and therapy. During the past decades an increased awareness is observed with respect to health related risks associated with ionizing radiation, both for patients as well as for medical personnel. As ionizing radiation is increasingly used, e.g. for imaging and diagnosis, there is a growing concern that the resulting exposure of patients and medical personnel may cause problems in the longer term. In particular, a link has been established between the accumulated dose of ionizing radiation received by a person and its risks of developing malignancies in the long term. As a precautionary measure, steps are now undertaken to impose guidelines and regulations on hospitals and clinical practitioners to keep record of the radiation doses received by a patient during all diagnostic and therapeutic procedures (the Radiation Passport). Hence, the availability of means to easily and correctly quantify the amount of radiation received during each procedure is of great importance.
In diagnostic and therapeutic procedures, skin is identified as one of the organs at risk in a broad subgroup of patients. The likelihood as well as the severity of radiation-induced skin effects increases with exposure dose. A dose of 2Gy (Gray) has been determined as the threshold above which deterministic (i.e. short term) effects will appear. However, there is no known threshold for the induction of skin cancer as a result of exposure to radiation.
Similarly, exposure of the skin to UV radiation can also lead to health problems.
Depending on the wavelength of the UV radiation, its effect will in fact be more or less harmful. Liv radiation is classified in three wavelength categories, respectively identified as A (315-400 nm), B (280-315 nm) and C (100-280 nm). While UV C is mainly blocked by the atmosphere and ozone layer, the population is exposed to UV A and UV B from the sunlight.
As it does not cause reddening of the skin due to (over)exposure, UV A was in the past considered to be the least harmful and the purpose of sunscreens was therefore limited to prevent UV B light to reach the skin. However, nowadays, the contribution of both UV A and UV B to skin cancer is recognized and recent sunscreens aim at protecting against both types of UV radiation. Depending on the sensitivity of their skin type, individuals are more or less sensitive to the deleterious effects of UV radiation. Infants are in general considered the most sensitive and vulnerable category.
To summarize, both in medical and non-medical situations, quantification of skin exposure to radiation (UV, X-rays etc.) is considered an important tool for health protection and for preventing future complications. To this end, the measurement of the radiation exposure on single spots or on more or less extended 2D surfaces is fundamental.
When considering the medical field, different techniques can be used to measure skin dose and determine the correlation with Peak Skin Dose (PSD), which is considered to be the most relevant parameter in skin dosimetry. Both direct as well as indirect skin dosimetry systems have been proposed and are used.
Examples of indirect skin dose quantification methods in medical practice are based on radiation beam-on time, dose-area product and cumulative dose calculation (at the interventional reference point). System output parameters can be combined with patient position tracking and possibly patient-specific data (such as weight and height) to obtain an estimated skin dose map. As most of these parameters can be obtained and read in real-time, they provide a valuable means of monitoring and possibly adapting the medical procedure at hand. However, as no direct measurement of radiation exposure to the skin is performed, their main disadvantage resides in the fact that indirect dosimetry only offers an estimated skin dose, based on forward projection of system output, instead of actual skin dose measurements. This may be particularly problematic in situations where radiation is delivered in a flexible manner, e.g. during diagnostic procedures requiring continuous or repeated imaging, while the patient also might change position. Direct skin dosimetry is based on actual measurement of skin exposure. This can be achieved on one or more discrete points, using e.g. MOSFET-based dosimeters or thermo-luminescent detectors (TLD5), or on a continuous two-dimensional surface using for instance radiographic or radiochromic films.
Although the region of high skin exposure is typically known in advance, the exact location of maximal exposure is not. This severely limits the likelihood of assessing PSD (Peak Skin Dose) with single-point dosimeters.
To some extent, these disadvantages can be overcome by using film-based dosimeters, i.e. radiographic or radiochromic films. While radiographic films are only useful up to dose levels of approximately 2Gy (which is generally considered to be the threshold for PSD, before deterministic effects appear), radiochromic films are capable of quantifying larger dose ranges, enabling their deployment both in diagnostic (or interventional procedures) as well as in therapy. However, film dosimeters are not easily used directly on the skin. Indeed, such usage would imply cutting the film and taping it, as to avoid the presence of an air gap between the film and the skin. Such air gaps could introduce artefacts in the dose measurements. Moreover, films are rather rigid and can only be bended within a limited range, without compromising the thickness of the active layer. Consequently, their use in complex anatomical regions (nose, ears and breast) may be problematic. Although in- vivo use of the films is feasible and has been demonstrated, it involves complex and time-consuming manipulations, eventually impacting on the workflow in the clinical unit.
Moreover, their read-out procedure is typically based on the measurement of their optical density by means of a transmission densitometer (flat scanner). This procedure can be performed only after completion of the treatment. Consequently, they can be used for retrospective quality assurance of the procedure. As such they are not suited for real-time quantitative skin dose monitoring with the aim of adapting the procedure, if needed, in order to minimize unwanted side effects on the skin or to adjust the dose for optimal efficacy.
With respect to the monitoring of UV radiation in the environment a similar distinction between direct and indirect skin dose measurements can be made. Certain applications use meteorological data for the relevant geographic location (UV index) to indirectly estimate the UV radiation dose experienced from exposure to sunlight. However, depending on the exact location, the estimated intensity of radiation can deviate substantially from the actual exposure intensity, especially under variable weather conditions.
Some systems are currently commercially available for direct UV exposure monitoring. UV sensitive sensors have already been integrated in watches and bracelets.
These systems generate UV exposure warnings, display UV index information or communicate with external systems, such as srnartphones, to inform the user about UV radiation exposure or even consult him/her regarding suitable protection measures such as sunscreen protection factors etc. Besides the need for power supply (possible by solar energy), the main disadvantage of this type of active systems probably resides in their cost price, rendering them less suitable for use with children. The latter market is mainly covered by the use of bracelets or patches changing color upon UV exposure. However these devices are totally passive and do not comprise any form of smart algorithm aimed at improving exposure estimation and allowing to cast an objective warning to the user when measures are to be taken to limit exposure, e.g. by (re-) applying a sunscreen or stepping out of the sunshine.
Summary of the invention
Although the methods and systems disclosed in the prior art provide useful solutions for performing skin radiation dose assessment under certain conditions, there still exists a long-felt need for an improved system and method for efficiently measuring skin radiation dose in general and peak skin radiation dose in particular, during medical procedures, such as diagnosis, image guided interventions and radiotherapy treatment. Furthermore, in non-medical settings, there is still a need for UV radiation dosimeters that are low cost, comfortable to wear and easy as well as objective to read out for routine use by the population.
It is an object of the present invention to provide efficient systems and methods for direct (electromagnetic and particle) radiation quantification on a surface in general, and skin radiation quantification in particular. Thereby, measurements can be made during the procedure or exposure to radiation, in real-time.
The above objective is accomplished by the present invention.
The present invention provides a system for measuring a radiation dose received by a pre-determined part of an object's external surface, in general, and a subject's skin surface, in particular. The system comprises a) an entity, hereafter referred to as a patch, that comprises radiation sensitive regions, hereafter referred to as functional cells, which, under the influence of (electromagnetic and/or particle) radiation, undergo physical and/or chemical changes that are expressed as a response signal taking the form of a measureable and quantifiable alteration in absorbance of a part of the electromagnetic wave spectrum; (b) a camera system which captures the relevant part of the electromagnetic spectrum, emitted in the environment and reflected on the patch, in the form of a digital picture, and, as such, allows quantification of the radiation-modulated patch-light interaction properties; and (c) a software algorithm -running on a control unit -which converts the digital picture into an estimate of the dose the patch was exposed to on the basis of said response signal.
In the following paragraphs, the properties of the different constituting elements are further elaborated on. Depending on the different embodiments of the invention, these properties have to be adapted as to obtain the desired functionality. Before continuing this refined description, some terminology is introduced in the following paragraph.
For reasons of clarity and brevity, the following terminology will be used in the remainder of this text, and has been used in the previous paragraphs. The term radiation refers to various kinds of particles with mass equal to zero (photons) or bigger, including, for example, UV radiation and different forms of ionizing radiation like X-ray, gamma-ray, alpha, beta and other forms of particle beam irradiation. The term light will refer to the part of the electromagnetic spectrum for which the patch referred to in part (a) of this invention exhibits radiation dose induced modulation of absorbance properties. As such, (some of) the sensors of the camera system, referred to in part (b) of this invention, should be sensitive to (a part of) this spectrum. Hence, light does not necessarily refer to the part of the electromagnetic spectrum visible to the human eye but may, for example, partially comprise the UV or IR part of the spectrum. Similarly, the term co/or refers to the absorbance properties of matter w.r.t.
the relevant electromagnetic spectrum covered by the term light.
The patch, as referred in part (a) of this invention, is composed of one or more components. In a preferred embodiment, the patch takes the form of a two dimensionally extended, or sheet-like, entity. The term patch is inspired by this preferred embodiment, without intending to limit the invention to it. Similarly, without restricting the invention to this preferred embodiment, the constituting components of the patch will be referred to with the term layers, with the top and bottom layers referring to the layers exposed to the outside world and facing the surface to be monitored, respectively, upon patch deployment. Hence, the patch is a monolayer or multilayer object, preferably, but not necessarily, taking the form of a two dimensionally extended, sheet-like, structure. The different layers composing the patch do not necessarily cover the entire patch extent.
The patch comprises at least one functional cell, which is characterized by a radiation dose dependency of its light absorbance properties, i.e. its color. By convention, this dependency is -within statistical limits -homogeneous along the functional cell's extent. The dose dependency of a functional cell's color typically takes the form of a gradual change of light absorbance properties upon increasing irradiation dose levels. Typically the proportionality of patch color with dose (i.e. first order derivative of color w.r.t. dose) eventually converges to zero (i.e. the functional cell's color saturates). In order to serve the overall functionality of the patch, at least a part of the layers possibly on top of the functional cell(s) are sufficiently transparent for the electromagnetic wave spectrum corresponding to the color range the functional cell covers upon radiation, and for which the camera used is sensitive.
In order to fulfill the aforementioned functionality, each functional cell comprises at least one radiation sensitive layer which undergoes physical and/or chemical changes upon irradiation. Also, each functional cell comprises at least one layer which undergoes a modulation of its light absorbance properties, i.e. a color change, as a direct or indirect result of the radiation-induced physical and/or chemical changes of the radiation sensitive layer.
The sensitivity of each functional cell, along with its saturation level, can be modified by changing the compositions of its constituting layers to match to the particular radiation dose levels and kind of radiation to be measured, as well as the properties of the camera system and SW algorithm used, as referred to in part (b) and (c) of this invention respectively.
Preferably, each functional cell is a monolayer entity, which comprises a radiation sensitive dye, which is immobilized in a solid yet deformable matrix (so as to maintain spatial information when a 2D assessment is needed). This could e.g. be achieved by the incorporation of radio-sensitive materials, such as, without being limited to, 10-12, pentacosadiynoc acid (PCDA) , leuco malachite green (LMG), turnball blue (TB), leuco crystal violet (LCV) or Diyne PC, in a gel or soft polymer matrix. The matrix may possibly further contain free radical initiators and/or scavengers or some dopant enhancing the color response of the patch. Upon irradiation the radio-sensitive materials will undergo radiation dose-related reactions like, for example, polymerization or oxidation reactions. These chemical reactions modify the physicochemical characteristics of the active components, which is macroscopically expressed as a modulation of its light absorbance properties, i.e. color change. Hence, in a preferred embodiment, the radiation sensitive and color changing layers coincide.
In an alternative embodiment, functional cells can be multilayer structures comprising a coinciding radiation sensitive and color changing layer, in combination with additional layers, e.g. to filter incoming and reflected light, or to transfer color changes to different spectral bands. Different multilayer functional cells on the patch can possibly share one or more of their constituting layers.
As an alternative embodiment, functional cells can be multilayer structures comprising a distinct radiation sensitive and color changing layer. Multilayer functional cells are possibly complemented with one or more intermediate layers to transfer the radiation induced physical and/or chemical changes of the radiation sensitive layer to a modulation of light absorbance of another layer. In addition, multilayer functional cells are possibly complemented with additional layers, e.g. to filter incoming and reflected light, or to transfer color changes to different spectral bands. Different multilayer functional cells on the patch can possibly share one or more of their constituting layers.
In a preferred embodiment, the patch is composed of one or more functional cells, exhibiting -within statistical limits -an identical dose dependent electromagnetic wave absorbance modulation characteristic.
In another preferred embodiment, the patch is composed of multiple functional cell groups. Each group contains a predefined configuration of functional cells, each having its own sensitivity and saturation level. This could be useful to increase the accuracy and/or robustness of the dose extraction, but also to measure different types of irradiation sim ultaneously.
The patch in its entirety must, upon application, fit with the shape of the body to which it is attached. This can in principle be achieved by pre-forming the patch as to fit onto the curvature of the body side to which it is attached. In case the size of the patch is small and it is to be attached to a relatively flat surface a flat patch may be adequate. In general, and in order to provide optimal comfort of use in various circumstances, it is preferred that the patch exhibits sufficient flexibility to deform along the curvatures of the surface to cover the relevant parts where irradiation is to be expected and measured. Preferably, the flexibility required is obtained by using material compositions for the patch layer(s) that are deformable themselves. In a less preferred embodiment, the patch is composed of multiple, possibly interconnected, rigid parts, or a composition of rigid parts on a deformable support or in a deformable matrix. The size of the composing rigid parts is such that they still allow the flexibility required to fit the patch on the curvature of the surface part to be monitored.
In particular cases such as e.g. point-like dose measurements, or 2D dose maps on flat surfaces (w.r.t. the extent of the monitored region), a rigid patch design can provide sufficiency of flexibility.
In a preferred embodiment of the invention, the bottom layer will be an adhesive layer which sticks onto the surface part to be monitored so that it will (1) keep the patch in a fixed position on the surface and (2) provide contact between patch and surface, reducing the likelihood of radiation dose influencing air layers to penetrate between patch and surface. In the particular case of skin dose measurements) the adhesive layer sticks the patch to the skin part to be monitored. Occasionally, the patch may be stuck onto a surface the exposure to radiation of which is representative for the exposure of the relevant area of the body, e.g. as piece of clothing surrounded by exposed skin. In another embodiment a part of the patch could be incorporated in pieces of clothing while another part -the dose-sensitive part -is, upon patch deployment, stuck onto the patch part incorporated into the cloths using e.g. a hook and loop fastener system (Velcro© type).
In a preferred embodiment of the invention, the functional cells are comprised in the top layers of the patch. In another preferred embodiment of the invention, the functional cells are covered by one or more) preferably inert) preferably radiation transparent and partially light transparent layers. Such layer can be protective) which may be warranted in order to improve the physical and/or chemical integrity, e.g. stability) of the system, for instance with respect to the presence of water.
The patch may be, but is not necessarily, enriched with any possible calibration information, facilitating the quantification of the light absorbance modulation and/or relating the latter to exposure to UV or ionizing radiation doses. Said calibration information may, for example, take the form of a set of delineated reference fields, the colors of which are stable and representative for a certain predetermined set of irradiation doses.
The light absorbance properties of the patch are captured and quantified using at least one reflectance camera system. Thereto) each camera involved comprises a set of light sensors. The ensemble of light sensors involved, possibly distributed over different cameras, is chosen such that the resulting spectral sensitivity at least partially matches the dose dependent color range covered by the functional cells incorporated in the patch. Hence, the camera(s) capture(s) the light, present in the environment and reflected on the patch. If the relevant electromagnetic spectrum is not or insufficiently present in the scene, it may be actively incorporated by the introduction of appropriate light sources.
Each camera involved generates at least one digital image. In this text, the term digital fmage refers to the digitized signal array generated by the camera sensor array, or any derived representation of it. Hence, depending on the camera and image representation considered a pixel value takes the form of a number or a higher dimensional tuple.
The signal generated by a sensor in a reflectance camera is proportional to the spectral integral of the incoming light energy and sensor (possibly with optical filter) sensitivity. Regarding sensors capturing light reflected from the patch, the spectral composition and strength of this incoming light is a complex interplay between environmental light colors and directions, as well as the imaged object's shape, texture (roughness) and its color. In the context of this patent application, all factors determining a pixel's value, except for the color of the object represented in that pixel, are referred to as confounding factors.
If needed, the data acquisition mode of the camera is attuned to the requirements imposed by the software algorithm used for conversion of the digital pictures generated by the camera to dose estimates.
Preferably, the dose dependent color range of the patch at least partially matches the visible light spectrum, enabling the use of standard and commonly available camera systems.
However, in an alternative embodiment, the patch may be chosen to react e.g. in the ultraviolet or infrared spectral domain, requiring specially adapted camera systems. The camera may be a stand-alone device (reflex) point and shoot...) or be part of another device (computer, smartphone, tablet...).
The software algorithm comprised in the invention, converts the digital image(s) acquired by the camera system(s) into a dose estimate. The latter can take the form of a single dose number (akin point measurements)) or a 2D dose map, possibly with residual measures of uncertainty. Depending on the particular patch design and acquisition mode used, the software algorithm can follow different approaches to extract dose information from images. Or, vice versa, depending on the software algorithm used, the patch design and data acquisition are attuned to allow dose extraction. These approaches mainly differ in the methodology chosen to separate the contribution to image pixel values of patch color from confounding factors (like lighting configuration and object shape), and to relate this to exposure dose. Typically, this dose extraction is performed following two steps, which are not necessarily clearly separated. In the first step, the pixel values of the digital image data are interpreted, which implicitly or explicitly amounts, amongst others, to separate object color information from confounding imaging factors. In the second step, interpreted pixel values are transformed into dose estimates by relating them to dose-color calibration. Dose-color calibration is first described in the following paragraphs before discussing particular embodiments of the software algorithms.
Transformation of interpreted pixel values to dose estimates requires some form of calibration relating the exposure doses to absorbance properties of the patch. This dose-color calibrotion can take several forms. Generally, dose-color calibration requires both dose measurement and, mostly, color measurement. Exposure doses are measured using direct or indirect dosimetry techniques, examples of which are already described in the introduction.
Color measurement and representation can take different forms, co-determined by the software routine used to convert patch color to dose information. Generally colors are represented using an ordered set of numbers. Well-known examples are the three-dimensional color representations, such as RGB, HSV, etc., but in the context of this text colors are represented in general as N-dimensional tuples, with N any integer number. These tuples are referred to as color coordinates in a particular co/or space. Hence, dose-color calibration relates color coordinates, or a property of their dose-dependent evolution, to dose measurements.
In a preferred embodiment, color measurement is performed using one or more reflectance cameras, with an ensemble sensitivity covering (a part of) the relevant electromagnetic spectral range. Acquiring image data of the patch, or at least its functional cells, upon radiation exposure, yields a set of calibration images for which corresponding exposure doses are measured, or could be derived. As previously mentioned, patch color is encoded in the particular image pixel values, along with confounding factors such as light conditions and object shape. In order to obtain useful dose-color calibration from image pixel values, corresponding to the patch's dose-sensitive areas, the confounding factors or at least differences in confounding factors within and between calibration images have to be, implicitly or explicitly, accounted for. In a particular realization this is achieved using a standardized acquisition protocol involving flat, and, compared to residual lighting inhomogeneity, small patch surfaces, along with uniform lighting and a fixed camera position, as orthogonal as possible to the flat patch surface. In another realization, color calibration patterns, comprising a known configuration of reference colors, are included in the scene.
The color space the color calibration pattern's colors are described in serves as reference color space, relating the image pixel values from the different images, or different regions of the same image. In all image-based calibration settings, the spectral composition of the lighting is chosen to, at least partially, match the dose-dependent color evolution of the patch, as well as possible color calibration elements in the scene.
In another embodiment, color measurement for dose-color calibration purposes is performed using optical density measurements of the patch's constituting components. This is particularly useful if the patch is designed as to incorporate dose-color calibration information. In this case, the patch colors corresponding to particular exposure doses have to be reproduced using a stable, dose-insensitive dye or, more general, material.
In a particular embodiment, and irrespective of the particular color measurement and color representation used, the dose-color calibration takes the form of a dose value for which the functional cell's color change saturates, possibly along with a representation of this saturated color. In another, more preferred, embodiment of the invention, dose calibration represents the evolution of the absorbance properties with exposure doses, irrespective of the particular color measurement and color representation used. This results in a -typically discrete -set of dose levels with corresponding color representation. The sampling rate of the exposure range is attuned to the patch sensitivity. Depending on the software algorithm used, and the sampling rate of the dose range, the calibration set is either retained as a discrete set or interpolated, or approximated, using a continuous representation such as linear interpolation or thin plate spline curve fitting. In an even more preferred embodiment, multiple color measurements are performed for each single exposure dose, enabling the construction of a statistical distribution describing the dose-dependent color evolution or derived values summarizing this distribution (such as median or mean).
Hence, in order to extract dose information from one or more images of the patch, the software algorithm relates image pixel values to dose-color calibration data.
In a first embodiment of the invention, the patch comprises one or more functional cells for which a dose distribution or dose average has to be determined based upon one or more images acquired after exposure (i.e. identical exposure represented in each image).
Hence, the software algorithm converts colors, opposed to color evolution or color change, to a dose distribution without explicitly requiring color information of multiple functional cells with different dose-dependent color properties. Hence, information on patch color or, more general, object color is extracted from image pixel values, and this information is related to a dose-color calibration. In this embodiment, the dose-color calibration takes the form of a set of color representations describing the patch's absorbance evolution over the relevant dose range, or a continuously approximated version of such discrete set. Different patch designs, algorithms and methods are suitable to realize this first embodiment. Depending on the approach taken, detection of the patch and, more specific, of its dose sensitive regions is implicitly or explicitly performed.
In a preferred realization of this first embodiment, contribution of patch color to image pixel values is separated from confounding factors by augmenting the scene with at least one color calibration pattern comprising a known configuration of colors with corresponding representation in a reference color space. In an even more preferred embodiment this color calibration pattern is integrated in the patch design. The reference color space is mathematically related to corresponding pixel values through the introduction of an appropriated parameterized imaging or shading model. Provided spatial correspondences between the color calibration pattern and the image values have been established, the unknown parameter values of the imaging model used are estimated.
Estimating the parameters of this shading model enables mapping of the reference color space to image pixel values and/or to transform image pixel values to the reference color space. Provided a known relation is established between the reference color spaces within which the color calibration pattern and the dose-color calibration colors are described, the imaging model allows to translate dose-color information to dose-image pixel values.
Preferably, the reference color spaces of the color calibration and dose-color calibration patterns are coinciding. In this embodiment the relation between dose-color calibration and image pixel values is established through the introduction of at least one color calibration pattern. As such, dose-color calibration information is not required to be explicitly included in the scene. The category this embodiment belongs to is therefore referred to as offline calibration.
In a very particular realization, the imaging model describes the patch and other relevant scene structures (such as the color calibration) as Lambertian surfaces upon exposure to multiple light sources, while the camera sensitivity is modeled with an affine matrix mixing the light and shape modulated object appearance into pixel values. A multivariate Gaussian noise model is introduced to describe the likelihood of observing a pixel value, given the corresponding object color, represented in the reference color space, and the shading model's parameters. Using pixel values corresponding to the color calibration pattern(s) and their known corresponding reference colors, aforementioned likelihood is maximized w.r.t. the unknown parameters of the shading model as well as the parameters of the Gaussian mixture model. Typically, shading models involve 3D information in the form of normal vectors. In a specific embodiment, this information is handled as such.
This is especially useful provided multiple images of the patch are available. However, in an alternative, simplified embodiment, this information is encoded into one or more mathematical functions extending over the 2D image. In an alternative embodiment) the Gaussian noise model is replaced by another statistical model providing better representation of the actual noise properties (e.g. Poisson noise model). Using the shading model) with its optimal parameter estimates, and the noise model as well as the particular (off/The) dose-color calibration information) the likelihood of observing a particular image pixel value after the corresponding patch region having been exposed to a particular radiation dose is formulated and maximized. When needed, this statistical model is extended with prior information, e.g. to regularize dose estimation results, in order to reduce the search space. In a particular realization, aforementioned statistical model is augmented to simultaneously detect dose sensitive regions.
Aforementioned embodiment relies on correspondences between image pixels and particular scene entities, such as color calibration regions. In a particular realization, these correspondences are established before color calibration, by spatial alignment) or registration, of the layout of the scene entities considered and image information. Such registration process preferably relies on information which has no or low sensitivity w.r.t. e.g. lighting conditions. In a very particular realization, this is achieved by the alignment of edges of the layout considered with gradient information extracted from the image(s) observed. In another very particular realization) dealing with vector valued pixels, this is achieved by alignment of normalized pixel values (e.g. pixel values divided by a vector norm (such as Li or L2)). In another very particular realization, a combination of the two aforementioned registration approaches is followed.
In a more advanced realization, the spatial alignment is performed simultaneously with the estimation of the shading model's parameters and the noise model. In an even more advanced realization, spatial alignment, estimation of shading model parameters, along with the noise model is performed simultaneously with the dose estimation.
In an alternative realization of this first embodiment, the relation between image pixel values and exposure doses is established through the introduction of at least one dose-color calibration region in the scene. In a preferred embodiment, this dose-color calibration pattern in included in the patch design. Opposed to the previously discussed off/me calibration approach, methods requiring dose-color calibration information to be included in the scene are referred to as online calibration approaches. As a dose-color calibration pattern can be interpreted as a color calibration pattern also, the previously described approaches are applicable. However, typically the dose-related color range does not form an optimal sampling of the color space, possibly introducing instabilities in the algorithm. As an alternative approach, methods for color constancy within an image are used to relate image pixel values to dose colors. Such approaches intend to map image pixel values to a new representation for which pixels corresponding to equally colored objects are mapped to the same value, while preserving differences between differing colors. Hence) by re-mapping the image pixels these approaches intend to remove confounding imaging factors (such as light and object shape) from the object's color information. After mapping of the (relevant) image pixels using the color constancy transform, the transformed dose-color calibration image pixels are used to infer dose estimates on the other image pixels in general, and on image pixels representing dose-sensitive patch regions, in particular.
In a particular approach of this embodiment, inspired by the work of Finlayson (Finlayson, Graham 0., Mark S. Drew, and Cheng Lu. "intrinsic images by entropy minimization." Computer Vision-ECCV 2004. Springer Berlin Heidelberg, 2004. 582-595.) Lambertian shading with Planckian lighting and narrow band camera sensors are assumed, and the resulting mathematical model is reformulated in order to allow removal of the confounding imaging factors. An important difference with the previously described approach, is that the color mapping applied does not relate image pixel values with a pre-defined reference color space. It only intends to remove the influence of confounding factors, such as light properties and object shape, from pixel values. However, as dose-color calibration information is incorporated in the scene, the dose-color calibration information is also subject to this mapping, rendering the use of a reference color space by the software algorithm unnecessary. In a particular realization, determination of this color constancy mapping is facilitated by the explicit use of the dose-color calibration regions, and possibly other color calibration regions present in the scene. In this case, those regions need to be detected before application of the color constancy approach. The registration approaches described in the previous embodiment are considered particular realizations for this registration task also. After application of the color constancy mapping, similar methods as previously described are used to relate dose-color calibration with the rest of the image pixels in general, and the image pixels representing the object's dose sensitive regions, in particular.
A very particular embodiment of the invention is feasible in particular settings in which dose-color calibration is introduced in the scene and in which the patch is relatively small compared to the surface curvature (i.e. patch shape can be considered to be flat), and lighting conditions can be considered uniform. In this case, the color constancy mapping or color normalization used in previous embodiments to separate object color from confounding factors is omitted. As such dose estimation amounts to direct comparison of image pixel values corresponding to dose-color calibration regions with image pixel values of the rest of the image in general, and the dose sensitive areas in particular.
In addition to the use of one or more images of identical exposure, as previously discussed, alternative embodiments of the invention exploit images taken during the exposure. These algorithms are provided with a set of images, each set corresponding to a particular exposure and containing one or more images. The previously described embodiments can be applied on each of these sets independently. However, extended embodiments exploit the time component inherently provided in the data to increase accuracy and/or robustness of the final dose estimates. In a particular realization, prior knowledge is introduced enforcing rnonotonically non-decreasing exposure doses. In other embodiments, a dose tracking approach is followed using Extended Kalman filters or particle filters or alike. Besides dose-related color, such approaches also exploit dose-related color evolution.
All aforementioned embodiments are capable of performing dose estimation using one functional cell. However) they are not limited to this scenario, and are also applicable when multiple functional cells are present in the patch. While, such scenarios require appropriate adaption of the dose-color calibration provided, they may increase robustness and/or accuracy of the dose estimates obtained.
In a second embodiment of the invention, the software algorithm infers dose estimates by comparison of dose-dependent colors of more than one functional cell having different dose sensitivities.
In a particular realization of this embodiment, the patchs dose-sensitive region is composed of measurement entities comprising a configuration of functional cells, each functional cell having its own dose sensitivity and saturation level. Upon exposure to radiation, the functional cells in a measurement entity change color according to their sensitivity giving rise to a color pattern. Assuming uniform exposure of a measurement entity, interpretation of this color pattern yields a dose estimate. In a very particular realization, dose estimation is based upon distinction between saturated and non-saturated functional cells of a measurement entity, and relating this distinction to known saturation levels of each functional cell. In another embodiment, super-resolution techniques are used to relax the assumption of uniform exposure. Aforementioned embodiments can also be used to complement the first embodiments of this invention, i.e. to relate the color pattern to known dose-calibration curves. This is beneficial to increase the sensitivity of the overall patch, at the cost of resolution.
The invention is further explained in the following examples which are intended to illustrate and not to limit the scope of the invention thereto.
Example 1
An example of patch design (template) is depicted in Fig. 1 and comprises a measurement area or (radiation sensitive) functional cell (1) and dose-color calibration regions (3). Each dose-color calibration region is comprised of six homogeneous colors, corresponding to six different doses (e.g. OGy, 0.5Gy, lGy, 1.SGy, 3Gy, 5Gy). In addition, a color constancy calibration region (2) is included in the patch design.
Figure 2 shows an image of the patch upon exposure to a radiation dose of 2Gy, obtained with a DSLR (reflex) camera. The functional cell (1) changed color as a consequence of the radiation exposure. Light variations in the patch will also be due to the curved shape of the patch and the shadows casted by the other structures in the scene. The image also comprises an arbitrary background (4) which is not used during the dose estimation.
The following approach was followed to obtain a dose estimate for this setup: * Image Preprocessing: The RAW image data extracted from the camera were converted into a RGB image. Hence, pixel values were represented as 3-tuples.
* Annotation of the patch areas: The different constituting components of the patch are labeled. In this very preliminary example, the annotation was performed manually. Hence, functional cell (1), dose-color calibration areas (3) and color constancy calibration region (2) were -roughly -indicated by the user. As an alternative, the patch template design could be registered to the images acquired to delineate the regions of interest.
* Color constancy: The input image pixel values were mapped to a derived representation in order to remove confounding influences of patch shape and light properties. Inspired by the work of Finlayson (Finlayson, Graham 0., Mark S. Drew, and Cheng Lu. "Intrinsic images by entropy minimization." Computer Vision-ECCV 2004. Springer Berlin Heidelberg, 2004.
582-595.), the patch was modeled as a Lambertian reflector, upon exposure to Planckian light and captured using a camera with narrow-band sensors. To obtain the particular color constancy transform, the color constancy calibration regions (2) were used in a two-step approach: o shape normalization: Using aforementioned assumption, shape information contributing to image pixel values is normalized for by conversion to normalized log RGB values. Hence, each image pixel value 3-tuple, is transformed with a logarithmic function. In this log RGB space, the contribution of shape to the transformed color coordinates is additive and identical for each color dimension. Hence, shape information can be removed by subtracting the log RGB coordinates with a (any) weighted average of themselves. The result of this operation is the mapping of the log RGB values to a two-dimensional plane, embedded in three dimensional log RGB color space. The resulting pixel value representations on this plane are referred to as log chromaticities. Obviously, depending on the choice of averaging coefficients, different log chromaticity representations are possible. Thereto, these averaging coefficients are chosen along a direction which is expected to carry little useful information for dose estimation purposes: the direction of the Planckian light variation in log RGB space. Indeed, possible dose-modulated patch color variations along this direction, cannot (at least without introduction of prior knowledge) be distinguished from light variations. This Planckian light variation direction is estimated using the principal component of the log RGB values corresponding to color constancy calibration regions. Once this principal component is determined, its coordinates are normalized as to obtain a set of averaging weights (having unity sum).
a Removal of residual light variation: Although the averaging weights in the previous step were chosen to be parallel to an estimate of the Planckian light variation direction in log RGB space, this weighted averaging does generally not amount to an orthogonal projection of log RGB values on a plane perpendicular to the Planckian light direction. As a result, light effects are still contributing to the log chromaticities values obtained in the previous step. According to the Planckian light model, this contribution, again, is oriented along one direction, with an amplitude (or inverse temperature) varying between different pixels. Hence, estimation of this direction, along with pixel-dependent Planckian light temperature variations, allows to remove the effects of light variation in the log chromaticities.
Again, Planckian light direction is estimated using the first principal component of the log chromaticities. As Planckian light temperature typically varies in a spatially coherent way (i.e. introduction of prior knowledge), the temperature values are modeled using a higher order polynomial defined on the image plane. The parameters of this smooth function are estimated by minimizing the variance of log chromaticities of the color constancy calibration regions after correction for the estimated Planckian light variation. Afterwards, this correction function is applied to all pixels in the image. The resulting image is referred to as normalized image.
* Dose-Color calibration: The dose-color calibration regions (3) in the normalized image, hence, after color constancy transform, are used to obtain a dose-color calibration curve. Specifically, an additive Gaussian noise model is introduced to model the normalized pixel value distribution corresponding to the six dose levels incorporated in the patch. The parameters of these distributions are obtained using Maximum Likelihood parameter estimation.
The normalized pixel value distributions corresponding to intermediate dose levels are, in this example, obtained by linear interpolation.
* Dose Estimation: A maximum likelihood approach is followed to obtain an estimate of the dose map. Thereto, a dose sweep is performed for each pixel.
Hence, for a pre-defined set of doses, the likelihood of each normalized pixel value is evaluated, and the dose corresponding to the highest likelihood is retained.
Example 2
In this example a different patch was designed. This is shown in figure 3. The patch comprises one circular functional cell (5), surrounded by a 15 color calibration region (6). The functional cell consisted of a flexible matrix, mixed with a dose-sensitive dye, PCDA. The patch was exposed to UV light upon which the functional cell (5) changes color. A picture was acquired using a DSLR (reflex) camera. A dose estimate was obtained using following procedure: * Image Preprocessing: The RAW image data extracted from the camera were converted into a RGB image. Hence, pixel values were represented as 3-tuples.
* Registration: The known patch design was used to automatically detect the patch and its constituting regions in an image following a three step procedure: o Initialization: An initial patch detection was obtained using a template matching approach. Thereto, a set of binary edge maps, corresponding to different patch sizes was constructed. The input image was convolved with a derivative of Gaussian kernel, to obtain smooth derivative images. A template matching was performed between each binary image on the one hand, and the norm of the derivative input image on the other hand. For every possible patch position in the image, and patch size considered, this yields a quality score identical to the sum of image gradient amplitudes in pixels coinciding with patch edges, as determined by the position and scale considered. The solution with the highest template matching score, normalized with the number of edge points in the template considered, was selected as initial patch pose and scale estimate.
o Iterative registration: The initial pose estimate was refined using an iterative edge-to-image-gradient registration approach. Thereto, a statistical model was formulated using -unknown, and unobserved -correspondences between template edge points on the one hand, and image pixels on the other hand. Provided that these correspondences are known, an aligning 2D/2D transform can be determined to match the patch with the image. The statistical model takes the form of a multivariate Gaussian and expresses the likelihood of observing a gradient vector in a particular image point, given that it would have been generated by a particular patch edge point. This likelihood decreases as the distance between gradient and edge point increases and as the orientation difference between the normal on the edge point and the image gradient direction increases. The prior likelihood to observe a correspondence between an edge point and an image point, is modeled proportional to the image gradient amplitude. An expectation maximization algorithm is used to obtain a maximum likelihood estimate of the registration transform parameters, relating the patch edges with the underlying image information. The transformation model considered in this examples is a 2D affine transform.
o Final circular registration: As the patch design is circular, containing a central circular functional cell (5), surrounded by 15 identically shaped color calibration regions (6), and edge/gradient information was previously used to align the patch with the image, there might be a residual rotational misalignment between patch and image. Therefore, the pixel values, representing color information, are used in this step to remove this rotational misalignment. However, as the colors between the patch layout and the image are substantially different (due to lighting, camera sensitivity, etc.) normalized RGB values are used to perform this final registration stage.
Thereto, the RGB values of each image pixel value, as well as each patch layout pixel, are divided by the sum of RGB values of that particular pixel. The resulting representation is less sensitive to confounding imaging factors. An exhaustive search comprising the 15 possible rotational residual misalignments is performed with a sum-of-squared different dissimilarity measure between the normalized patch layout and the normalized image RGB values.
* Color matching: After having established correspondences between patch layout and image pixels, the influence of confounding factors to the input image's pixel values was removed. Thereto, a very simple imaging model was formulated, assuming the patch to be a flat, Lambertian reflector upon exposure to uniform (ambient) lighting.
The camera sensitivity was modeled using an affine mixing matrix. As a result, an affine color transform relates RGB pixel values of the image with reference RGB pixel values of the patch layout. A multivariate Gaussian noise model was formulated to express the likelihood of observing an image pixel value, given the color transformed value of its corresponding patch layout pixel. A maximum likelihood approach was used to estimate the parameters of this affine color transform, as well as the parameters of the noise model.
* Dose Estimation: A dose-color calibration curve, relating reference RGB pixel values of the patch with corresponding exposure doses, was used for dose estimation purposes. The construction of the dose-color calibration curve is discussed later.
Similar to example 1, a maximum likelihood dose estimate is obtained by performing a dose sweep. The color matching transform determined in the previous step is used to transform the reference RGB values corresponding to the different dose levels considered to the RGB space of the input image. Hence, in this example off/me calibration is used. The Gaussian noise model, introduced and determined in the previous step, is used to evaluate the likelihood of observing a particular image pixel RGB value given the transformed reference RGB values corresponding to the dose levels considered.
The dose color-calibration curve used in this example was obtained as follows: * Data Acquisition: A patch -identical to the one used for dose estimation purposes -was incrementally exposed to UV light. After each incremental exposure, an image was acquired using a DSLR (reflex) camera.
* Registration: The patch layout was aligned with each image obtained using previously described method.
* Color Matching: An affine color matching transform relating reference RGB pixel values to image RGB pixel values was determined for every calibration image, using previously described method.
Dose-Color Calibration Curve: The previously determined affine color matching transforms are used to transform calibration image pixel values corresponding to the patch's functional cell (5) to the reference RGB color space. For every dose level considered, the median of these transformed image RGB pixel values is determined and used for calibration purposes. A thin plate spline curve, fitting the calibration data, was determined to establish a continuous relation between dose levels and reference RGB pixel values.