Classifications

• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T5/00—Image enhancement or restoration
  • G06T5/70—Denoising; Smoothing
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/10—Image acquisition modality
  • G06T2207/10116—X-ray image
  • G06T2207/10121—Fluoroscopy
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/20—Special algorithmic details
  • G06T2207/20004—Adaptive image processing
  • G06T2207/20012—Locally adaptive
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/20—Special algorithmic details
  • G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/20—Special algorithmic details
  • G06T2207/20172—Image enhancement details
  • G06T2207/20182—Noise reduction or smoothing in the temporal domain; Spatio-temporal filtering
• • G—PHYSICS
  • G06—COMPUTING OR CALCULATING; COUNTING
  • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
  • G06T2207/00—Indexing scheme for image analysis or image enhancement
  • G06T2207/30—Subject of image; Context of image processing
  • G06T2207/30004—Biomedical image processing

Definitions

• the invention is in the field of X-ray imaging and, more specifically, that of medical fluoroscopy imaging. It relates to a method of reducing noise in a sequence of fluoroscopic images acquired by an X-ray detector.
• X-ray fluoroscopy imaging makes it possible to provide a flow of images from a patient to a physician during so-called minimally invasive surgical procedures, for example the chemoembolization of the liver. vertebroplasty, catheterization of aneurysms or treatment of vascular stenosis.
• the images can notably help guide surgical instruments.
• a fluoroscopically guided intervention is typically performed by passing a catheter within the vascular network of the patient.
• a contrast agent may or may not have been previously injected into the vascular network in order to opacify it and improve the visibility of the vessels.
• Such an intervention is generally relatively long and the dose of X-rays to which the patient is subjected must be limited in order to avoid causing tissue lesions or burns. Because of this limitation of the X-ray dose used, the fluoroscopic images obtained have a relatively high noise level, and therefore a relatively low contrast-to-noise ratio, making the images difficult to read.
• Filtering treatments are implemented in order to reduce the quantum noise present in these images and increase their contrast-to-noise ratio.
• FNR treatment A treatment known under the name of fluoroscopic noise reduction treatment, or FNR treatment according to the abbreviation of the English expression "Fluoroscopy Noise Reduction”, is thus generally performed on the images acquired by the X-ray detector.
• the objective of this FNR treatment is to filter the noise present in the image while preserving the contrast of the information present in this image.
• the FNR is performed by applying a temporal filter to the regions of the images where there is no displacement. The existence or absence of displacements in regions of the image is detected from the individual variation of the intensity of each pixel considered separately.
• a pixel is considered to be in motion when its intensity variation between two images exceeds a threshold linked to the standard deviation of the noise.
• a recursive temporal filtering treatment is applied to the pixels determined to be fixed.
• the FNR treatment has the advantage of not coloring the noise spatially. In other words, it does not reveal fictitious objects resulting from groupings of pixels of intensities close to each other.
• the FNR treatment has a relatively limited denoising capacity and tends to remove objects of interest in the image.
• spatial filtering has interesting properties.
• filtering by image transformation in the wavelet domain is one of the most used techniques.
• the wavelet theory introduced in the 1980s, was developed from the Fourier transform.
• the Fourier transform is used to describe many physical phenomena. It is indeed a powerful mathematical tool to describe many physical phenomena, including physical phenomena modeled by stationary signals.
• the representations of physical phenomena modeled by non-stationary signals are not generally satisfactory.
• the Fourier transform is unable to locate the portions of the signal where the frequency suddenly changes.
• a Fourier transform with windows has been developed. It consists of multiplying the signal by a window of given dimension.
• the analysis thus becomes local, the size of the window determining the temporal resolution obtained, or the spatial resolution in the case of two-dimensional signals.
• the wavelet theory is based on the Fourier transform with windows. It differs in that it allows, during the transformation, to change the size of the window. For this reason, we speak of multiresolution analysis or multiscale analysis. Another difference is that the signal is no longer broken down into a sum sine and cosine, but according to functions called "wavelets".
• the wavelets have the advantage of transforming the image into a so-called hollow representation.
• hollow representation is meant a representation requiring a small number of parameters to faithfully represent the image.
• An object of the invention is notably to provide a denoising treatment adapted to fluoroscopic images whose contrast to noise ratio is relatively low.
• the denoising treatment must have a high denoising capacity, generate a low calculation cost, limit the loss of objects of interest in the image, and not introduce artifacts.
• the subject of the invention is a method for reducing noise in a sequence of fluoroscopic images acquired by an X-ray detector, each image being formed in the spatial domain of a matrix of pixels, each pixel having a value representative of a signal level, the method comprising the following successive steps for each image:
• ⁇ apply temporal filtering on the image acquired at an instant n, the application of said temporal filtering comprising the following substeps for each pixel of the image:
• ⁇ apply spatial filtering on the image acquired at time n, the application of said spatial filtering comprising the following substeps:
• thresholding the coefficients of the image by a thresholding function, the thresholding function canceling the coefficients lower than a third predetermined threshold, and keeping or adjusting the coefficients greater than the third predetermined threshold,
• the stage of transformation of the image acquired at time n in the field of curvelets advantageously uses a transform into curvelets with 6 or 9 scales, and / or 16 orientations.
• the step of transforming the image acquired at time n in the curvelet domain uses a discrete curvature transform by USFFT.
• the step of transforming the image acquired at time n in the curvelet domain uses a discrete curvelet transformation by wrapping.
• the thresholding function of the spatial filtering may be a hard thresholding function, that is to say a thresholding function according to which the coefficients are canceled if they are lower than the third predetermined threshold, and otherwise preserved.
• the third predetermined threshold T s of the spatial filtering can be determined by the following successive steps:
• ⁇ determine one of the coefficients c i Y k by the relation: where M and P designate the dimensions of the image Y and / denotes the orientations of the curvelets,
• ⁇ is the standard deviation of the noise in the image acquired at time n and on which the temporal filtering has been applied
• the factor r takes the value 0 for the coarsest scale of the curvelets, the value 4 for the finest scale, and 1 for the other scales.
• the step of applying the temporal filtering on the image acquired at the instant n comprises, following the substep of determining the difference between the value of the pixel considered in the image acquired at time n and the value of the corresponding pixel in the image acquired at time n-1 on which temporal filtering has been applied, the following additional sub-step:
• the noise reduction method according to the invention successively comprises temporal filtering and spatial filtering by curvelet transformation.
• the order in which the filtering is carried out is of importance with regard to the coloration of the noise.
• spatial filtering by curvelet transformation introduces artifacts if it is not preceded by temporal filtering.
• the invention has the particular advantage of not requiring a priori forms for spatial filtering.
• FIG. 1 represents an example of points taken into account for determining the contrast of an object in a fluoroscopic image
• FIG. 2 represents possible steps of the noise reduction method according to the invention
• FIGS. 3A and 3B show, by curves, a hard thresholding function and a soft thresholding function, respectively;
• FIG. 4 illustrates, by a graph, the performance of this noise reduction method for a sequence of fluoroscopic images representing an object having a translational and rotational movement
• FIG. 5 represents an example of sub-steps of the noise reduction method making it possible to perform spatial filtering
• FIG. 6 represents an example of a thresholding function used for the spatial filtering of FIG. 5.
• CNR contrast-to-noise ratio
• FIG. 1 represents the points considered (in white) for determining the contrast of a stent, better known by the English name of "stent".
• Figure 2 shows possible steps for the noise reduction method in the fluoroscopic image sequence. The method is described for the image x n acquired at time n. However, it can be applied successively to each of the images in the sequence.
• the image x n is processed by a preprocessing chain.
• Pretreatment may include groupings of adjacent pixels in the image.
• the pixels can be grouped for example according to the desired resolution or the desired noise level.
• the pretreatment may also comprise a phase of calibration of the X-ray detector. In particular, the gain and the offset to be applied to each pixel of the detector may be determined.
• the calibration may also include an identification of the pixels taking a random value or a constant value regardless of the received radiation. These defective pixels can be corrected by interpolating their value by the values of the neighboring pixels.
• the pretreatment may also include rescaling the images, i.e., an overall adjustment of the value of all pixels of each image, so that the average pixel value is constant for all footage of the sequence.
• the preprocessing may comprise a substep of adaptation of the images.
• the noise present in each image x n is a fish noise. Such noise is unsuitable for noise reduction treatments.
• each image x n is a representative image of X-ray attenuations, while an image of radiological thicknesses is more natural to analyze. Ascombe's formula makes it possible to obtain an image of the radiological thicknesses with a pseudo-Gaussian noise:
• x 0 n (m, p) is the value of the pixel (m, p) before the substep of adaptation of the image x n .
• the image x n is filtered by a temporal filter.
• the temporal filtering is for example a filtering called "Fluoroscopic Noise Reduction" (FNR). This filtering is described in particular in US Pat. No. 6,314,160 B1.
• FNR Fluoroscopic Noise Reduction
• a (m, p) the difference between the value x n (m, p) in the image x n acquired at time n and the value y n _i (m, p) in filtered image n. ! corresponding to the acquisition at time n-1:
• FNR filtering is based on two modes, called adaptive mode and non-adaptive mode.
• the adaptive mode is called because of its consideration of the movement of objects.
• FNR filtering is to compare pixel by pixel, the difference ⁇ at predetermined thresholds.
• the first threshold Si corresponds to a relatively small difference ⁇ .
• the second threshold S 2 corresponds to a relatively high difference ⁇ .
• FNR filtering is then parameterized by coefficients a and ⁇ as a function of the result of this difference ⁇ and the mode considered.
• the strength of the filter is:
• ⁇ DD ⁇ (1 - a)
• the thresholds Si and S 2 are defined as follows:
• the thresholds Si and S 2 are defined as follows:
• FNR filtering itself consists in determining a new value y n for each pixel of the image from its value x n and the coefficients a and ⁇ :
• y n a. x n + (1 - a) n _ ! + ⁇
• the filtering step FNR is shown for the adaptive mode.
• the absolute value of the difference ⁇ is compared for each pixel x n at the threshold S lt, namely the standard deviation o. If the absolute value of the difference ⁇ is less than the S lt threshold the pixel is probably a high noise level and said filter is applied in a sub-step 222 with the corresponding parameters. Otherwise, in a sub-step 223, the absolute value of the difference ⁇ is compared for each pixel x n at the threshold S 2 , namely 2o.
• the pixel considered corresponding probably to a point of the moving object and so-called weak filtering is applied in a substep 224 with the corresponding parameters.
• the absolute value of the difference ⁇ is between the thresholds 5 X and S 2 , an uncertainty exists.
• an average filtering is applied in a sub-step 225 with the corresponding parameters.
• a third step 23 the image y n obtained after time filtering is filtered by a spatial filter based on the curvelet transform.
• the theory of curvelets uses the principle of multiresolution or multiscale spaces, a scale corresponding to a partition level of space.
• the space is partitioned into curvelets whose envelopes are of dimensions 2 j x 2 d / 2 , where j is a positive integer denoting the scale.
• Each curvelet depends on the number of scales considered, its position in the image and its orientation. For each scale, the best possible approximation of the image is calculated. From one scale to another, only the improvements obtained are represented. The number of scales determines the quality of the reconstruction of the image.
• the image y n is transformed from the spatial domain to the curvelet domain by a transformation into curvelets.
• a discrete curvelet transformation is used.
• Such a transformation is notably described in E. Candès, Demanet L, Donoho D. and Ying L, "Fast Discrete Curvelet Transform", Multiscale Model. Simul., Vol. 5, no. 3, 2006, pp. 861 - 899.
• the curvelets can not be sampled according to the Fourier space grids.
• the curvelet transform of the image y n can be represented by a set of coefficients c (j, l, k), where the exponent D refers to the discrete transform, and where j, l and k denote respectively the scale, the orientation and the position of the curvelet considered.
• the number of scales used is equal to 9.
• the number of scales used is equal to 6. This last embodiment implies a smaller number of coefficients to calculate and therefore a lower processing time.
• the number of orientations used is advantageously equal to 16. This number presents a good compromise between the resolution in orientation and the computational cost of the curvelets, in particular for real images of which curves require a large number of orientations to approach them. Moreover, the number 16 being a multiple of 4, it facilitates the transformation into curvelets by wrapping.
• the coefficients c (j, l, k) of the image y n are thresholded by a thresholding function.
• the thresholding step 232 is based on the assumption that some coefficients c (j, l, k) represent the noise while others contain the useful information.
• the thresholding step 232 therefore consists of identifying and deleting the coefficients that do not contain useful information.
• the thresholding function is characterized on the one hand by its shape and on the other hand by its threshold T. There are mainly two thresholding functions, the so-called soft thresholding function and the so-called hard thresholding function. By simplifying the notation of the coefficients by c n and calling c n the thresholded coefficients, the hard thresholding function can be modeled as follows:
• the soft thresholding function can be modeled as follows: if not
• FIG. 3A shows a curve 31 the hard thresholding function and FIG. 3B shows a curve 32 the soft thresholding function.
• the soft thresholding function has a greater denoising capacity than the hard thresholding function.
• it modifies the value of the pixels according to the threshold T chosen. This results in a loss of precision on certain details of the image.
• the determination of the threshold T is a crucial step because it determines the performance of spatial filtering, both in terms of denoising capacity and the introduction of artifacts.
• the threshold T is determined as being the universal threshold ⁇ ⁇ described in particular in DL Donoho and Johnstone, IM, "Ideal spatial adaptation via wavelet shrinkage", Biometrica, vol. 81, 1994, pp. 425-455.
• This threshold ⁇ ⁇ is defined by the following relation:
• N is the number of samples, i.e. here the number of pixels in the image y n .
• the universal threshold ⁇ ⁇ is not optimal and can introduce artifacts. However, thresholds proportional to r universal threshold there may be good results if the proportionality factor is chosen.
• the threshold T is determined by the method described in JL Starck, Candès IJ and Donoho L, "The Curve Transform for Image Denoising", IEEE Transactions on Image Processing, vol. 1 1, no. 6, 2002, pp. 670-684.
• This method consists in creating an image, denoted by Y, of the same dimensions as the image y n to be denoised and of uniform value 1 on all the pixels.
• the Y image is transformed by the Fourier transform.
• the coordinate frequency (0,0) in the image Y is centered and its amplitude is normalized by the dimensions of the image.
• F be the amplitude of the central frequency. Its normalized amplitude, noted
• M and P designate the dimensions of the image Y, that is to say also of the image x n and of the image y n .
• T s The threshold of this method, denoted T s, depends on the scale j and on the orientation l. It is determined according to the standard of the coefficients c i Y k of the curvelets of the image Y for a scale j and a direction l data. This standard c i Y is determined by the following relation:
• the thresholding function is a hard thresholding function, defined as follows:
• the image y n is transformed from the domain of curvelets to the spatial domain (matrix) by a transformation into inverse curvelets.
• the threshold coefficients are used.
• the image obtained after temporal filtering and spatial filtering is denoted z n .
• FIG. 4 graphically illustrates the performance of the noise reduction method according to the invention for a sequence of images representing a stent having translational and rotational movement.
• the x-axis represents the average CNR of the x n sequence of images and the y-axis represents the average CNR of the image sequence obtained after noise reduction processing.
• a first curve 41 represents the average CNR of the image sequence z n obtained by the noise reduction method according to the invention, that is to say by a temporal filtering followed by a spatial filtering by transformation into curvelets.
• a second curve 42 represents the average CNR of the same sequence of images after processing by the same temporal filtering followed by spatial filtering by wavelet transformation.
• the graph shows that for a sequence of images to be processed with a relatively low CNR (0.5-0.7-1), the noise reduction method according to the invention shows a significant improvement in terms of CNR.
• the noise reduction is even better for a sequence of images with a higher CNR (1, 5 - 2 - 2.5).
• the noise reduction method according to the invention makes it possible to significantly reduce the introduction of artifacts.
• the spatial filtering step 23 described above makes it possible to effectively reduce the noise level in the images.
• this step has the disadvantage of introducing artifacts appearing in the form of curvelets, that is to say objects with elongated and undulating shapes. Indeed, these artifacts are due to coefficients having a large amplitude, but representing only noise. They are present only when the noise level is high, which is the case of fluoroscopic images.
• One solution is to set relatively low thresholds, that reduces artifacts but limits performance in terms of noise reduction.
• the following description describes another method for performing the spatial filtering step to significantly reduce the presence of artifacts without reducing the performance of the filtering.
• This method is based on the correlation that exists between a given coefficient and the coefficients that are related to it, namely its parent coefficient, its neighboring coefficients and its cousin coefficients.
• a coefficient represents the useful signal
• the coefficients that are bound to it have a high probability of also representing the useful signal.
• the modulus of a coefficient is large because of the presence of an artifact, it is unlikely that the coefficients related to it are also affected by this artifact.
• FIG. 5 represents another example of sub-steps for performing the spatial filtering step 23.
• the image y n obtained after temporal filtering is transformed from the spatial domain to the curvelet domain.
• Transform curvelets image y n is represented by rated ° coefficients c (j, l, k), or c n in simplified form.
• the transformation into curvelets is done with 6 scales and 12 orientations.
• T x , T 2 and T 3 are determined such that T l ⁇ T 2 ⁇ T 3 .
• These thresholds are determined according to the standard deviation a sfl of the subband of the coefficients.
• the subband of a given coefficient c ° (j, l, k) groups all the coefficients having the same scale j and the same orientation /. It is therefore the set of positions for the scale and orientation considered.
• the thresholds ⁇ ⁇ , T 2 and T 3 define a first zone for the values below the threshold ⁇ ⁇ , a second zone for the values between the thresholds T x and T 2 , a third zone for the values between the thresholds T 2 and T 3 , and a fourth zone for values above the threshold T 3 .
• the coefficients are corrected according to their parent coefficient and their neighboring coefficients. Of preferably, only the uncorrelated neighbor coefficients are considered.
• c ° (j, l, k), or c n in simplified form, the corrected coefficients.
• the second case is where the modulus of the parent coefficient c ° (jl, l, k) and the mean of the modules of the neighboring coefficients c ° (j, l, k ') lie in the same area, different from that of the module the coefficient considered c (/, /, &).
• the modulus of the corrected coefficient c ° (j, l, k) then takes the minimum value of the area in which the modulus of the parent coefficient is situated and the average of the modules of the neighboring coefficients if the modulus of the coefficient considered c ° (j, l, k) is in a zone with lower values, and the maximum value of this zone otherwise.
• the third case is the one where the modulus of the coefficient considered c ° (j, l, k), the modulus of its parent coefficient c ° (jl, l, k), and the mean of the modules of its neighboring coefficients c ° (j , l, k ') are in three distinct zones.
• the modulus of the corrected coefficient c (j, l, k) then takes the minimum value of the area in which the average of the modules of the neighboring coefficients is situated if the modulus of the considered coefficient c ° (j, l, k) lies in a zone with lower values, and the maximum value of this zone otherwise.
• the third case favors the zone of the uncorrelated neighboring coefficients to the detriment of the zone of the parent coefficient because at the coarsest scale, the coefficients do not have a parent. It would nevertheless be possible to correct the coefficients of other scales according to the area in which the module of their parent coefficient is located.
• the corrected coefficients c n are thresholded by a thresholding function.
• a thresholding function We write c ° (j, l, k), or c n in simplified form, the thresholded corrected coefficients.
• FIG. 6 represents the thresholding function corresponding to the threshold values T x , T 2 and T 3 considered. This function is modeled as follows:
• f ⁇ c n is a sigmoid defined as follows:
• the parameters m and ⁇ are determined so that the thresholding function is continuous. They are thus defined as follows:
• the thresholding function thus cancels the coefficients whose modulus is included in the first zone 61, it applies a soft thresholding to the coefficients whose module is included in the second zone 62, it applies a hard thresholding to the coefficients whose module is included in the fourth zone 64, and it applies a transition threshold to the coefficients whose module is included in the third zone 63.
• the soft thresholding makes it possible to reduce the noise level, but it spreads the outlines. Hard thresholding keeps the contours well but is less efficient in terms of noise reduction. The joint use of these two types of thresholding makes it possible to combine their respective advantages according to the zones.
• threshold T 4 is defined from the standard deviation a sfl of the subband of the coefficient c "considered:
• a fifth substep 55 the image y n is transformed from the curvelet domain to the spatial domain by a transformation into inverse curvelets.
• the corrected threshold coefficients are used.
• the image obtained after temporal filtering and spatial filtering is denoted z ' n .
• the spatial filtering described with reference to Figure 5 uses both the parent coefficients and the uncorrelated neighbor coefficients. Spatial filtering could also use correlated neighboring coefficients, or the set of neighboring coefficients. It could also use only the parent coefficients or the neighboring coefficients, or combine these coefficients with the cousin coefficients. These coefficients are those having the same scale and the same position as the reference coefficient.
• the spatial filtering could also not include a sub-step of correction of the coefficients, the sub-step 54 of thresholding being in this case carried out directly on the coefficients c n of the transforms in curvelets.

Landscapes

• Physics & Mathematics (AREA)
• General Physics & Mathematics (AREA)
• Engineering & Computer Science (AREA)
• Theoretical Computer Science (AREA)
• Apparatus For Radiation Diagnosis (AREA)
• Image Processing (AREA)

Priority Applications (3)

Applications Claiming Priority (2)

Publications (1)

Family

ID=46548470

Family Applications (1)

Country Status (5)

Cited By (1)

Families Citing this family (19)

Citations (4)

Family Cites Families (11)

• 2011
  • 2011-07-22 FR FR1102306A patent/FR2978273B1/fr not_active Expired - Fee Related
• 2012
  • 2012-07-20 JP JP2014520683A patent/JP6120843B2/ja active Active
  • 2012-07-20 US US14/234,292 patent/US9058654B2/en active Active
  • 2012-07-20 WO PCT/EP2012/064347 patent/WO2013014110A1/fr not_active Ceased
  • 2012-07-20 EP EP12737811.5A patent/EP2734975B1/fr active Active

Patent Citations (4)

Non-Patent Citations (4)

Also Published As

Similar Documents

Legal Events