Automatic segmentation of the inner, middle and outer ear

Abstract

The present invention relates to a method for automatically segmenting the human cochlea, comprising an image (M) composed of voxels (v), each voxel (v) including an intensity value indicative of the intensity of each voxel (v).

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of priority from earlier filed European patent application no. 22159818.8, filed March 2, 2022. [Technical field]

[0002] The present invention relates to a method for detecting and segmenting the human cochlea, and a corresponding system and computer program. [Background technology]

[0003] The inner ear (auris interna) is the innermost part of the human ear and consists of two important functional parts: the cochlea, which converts sound pressure from the outer ear into electrochemical impulses that are transmitted via the auditory nerve to the brain for sound perception, and the vestibular system, which acts as a means of maintaining balance. These structures are so small that surgery performed on these organs requires the utmost precision.

[0004] Typical surgeries performed on the human ear are acoustic nerve surgery to remove acoustic nerve tumors, specifically either a translabyrinth surgery in which the mastoid bone and bones of the inner ear are removed to gain access to the ear canal for removal of the tumor, a retrosigmoid sinus postcranial / suboccipital surgery in which the surgeon makes an incision through an opening in the skull behind the mastoid portion of the ear, a middle fossa surgery in which the surgeon removes tumors from the upper surface of the inner ear canal beyond the inner ear, and a transauricular surgery in which the mastoid bone and bones of the inner ear are removed to gain access to the ear canal for removal of the tumor.

[0005] Another representative surgery is cochlear implant (CI) surgery. Such surgically implanted electronic devices help restore sound perception to individuals with severe hearing loss, for example due to damage or defects in the inner ear. In particular, such cochlear implants can be configured to directly stimulate the auditory nerve to send information to the brain.

[0006] In particular, in the natural hearing process, the anatomy of the ear mechanically converts sound into vibrations of the basilar membrane of the cochlea, which stimulate nerves connected to the spiral ganglion and ultimately the auditory nerve, with higher frequencies stimulating the more basal spiral ganglion nerves, while lower frequencies stimulate the more distal spiral ganglion nerves. Thus, cochlear implants can be configured to exploit this so-called natural tonotopy to induce perceived lower (higher) frequency sounds by applying an electric field to the more distal (basal) spiral ganglion nerves.

[0007] In this way, the cochlear implant induces hearing by stimulating auditory nerve pathways in the cochlea, for example with an implanted electrode array. The cochlear implant's processor can be programmed to process sounds detected by the microphone and send signals to each electrode. The cochlear implant electrode array is configured such that, when properly positioned, each electrode stimulates a nerve pathway corresponding to a predetermined frequency bandwidth. However, in surgery, the cochlear implant electrode array is blindly passed into the cochlea, and its insertion path is guided only by the walls of the spiral cochlear cavity. Since the final location of the electrodes is generally unknown, the only option when programming the cochlear implant is to assume that the electrodes are optimally positioned in the cochlea and use a default frequency allocation table.

[0008] Yet another surgical procedure is ear reconstruction in congenital anotia, specifically aimed at reconstructing the external auditory canal, tympanic membrane and ossicular chain (hearing bones of the middle ear).

[0009] Additionally, by implanting a bone-integrated bone conduction hearing aid system, a bone-anchored hearing aid can be surgically implanted into the skull, transmitting sound through the bone to the inner ear.

[0010] Additionally, the stapes can be removed through a procedure called a stapedectomy / stapedotomy and replaced with an artificial bone.

[0011] Yet another surgical procedure, tympanoplasty, may be performed on the tympanic membrane and middle ear bones to restore the hearing mechanism of the middle ear.

[0012] All of the above implants / surgical procedures benefit greatly from precise knowledge of the position and shape of each part of the ear.

[0013] Therefore, the problem that the present invention aims to solve is to provide a method, system and computer program for reliably detecting and segmenting the human cochlea.

[0014] This problem is solved by a method having the features of claim 1, a system having the features of claim 17 and a computer program having the features of claim 18.

[0015] Preferred embodiments of these aspects of the invention are defined in the corresponding dependent claims and are described below.

[0016] According to claim 1, there is provided a method for automatically segmenting the human cochlea contained in an image M composed of voxels v, each voxel comprising an intensity value indicative of the intensity of the respective voxel v, said method comprising: - Voxel V E1 a first empty gradient accumulation image E1 constructed from the same origin and size as the image M, and each voxel v E1 to a voxel v of the image M, wherein for each voxel v of the image M, a gradient vector g pointing in the direction of the largest possible intensity increase of the image M and a first gradient line extending in the direction of the gradient vector g, the first gradient line starting from a first starting point p start,1 Starts at the first end point p end,1 and for each voxel v of the first gradient accumulation image E1 where a corresponding voxel v of the image M intersects with the respective first gradient line, E1 by a predefined constant increment; - Voxel V E2 a second gradient accumulation image E2 having the same origin and size as the image M, and each voxel v of the second gradient accumulation image E2 being E2 provides the second gradient accumulation image E2 associated with voxels v of the image M, wherein for each voxel v of the image M, a gradient vector g pointing in the direction of the largest possible intensity increase of the image M and further a second gradient line extending in the opposite direction of the gradient vector g, the second gradient line starting from a second starting point p start,2 Starts at the second end point p end,2 and for each voxel V of the second gradient accumulation image E2 where a corresponding voxel v of the image M intersects the respective second gradient line, E2 by a predefined constant increment; - Voxel V E3 a third gradient accumulation image E3 having the same origin and size as the image M, and each voxel v of the third gradient accumulation image E3 being E3 to a voxel v of the image M, wherein for each voxel v of the image M, a gradient vector g pointing in the direction of the largest possible intensity increase of the image M and a third gradient line extending in the direction of the gradient vector g, the third gradient line starting from a third starting point p start,3 Starts at the third end point p end,3 and for each voxel v of the first gradient accumulation image E3 where a corresponding voxel v of the image (M) intersects with the respective third gradient line, the third gradient line ending with E3 by a predefined constant increment; Using the first, second and third gradient accumulation images (E1, E2, E3), voxel v c A cochlear descriptor image E consisting of descriptor and the intensity values ​​are expressed as the cochlear descriptor image E descriptor is greater than the intensity values ​​of all other voxels in the cochlear descriptor image E descriptor Number of voxels v ≥ 1c and using this to predict the likely location of the center of the cochlea in the image M.

[0017] In particular, according to one embodiment of the present invention, a voxel v c The selection of the number P≧1 is based on the cochlear descriptor image E descriptor with a value obtained by multiplying the brightest voxel (i.e., highest intensity) of the descriptor image by a factor in the range of 0.6 to 0.8, preferably the factor is 0.7. This thresholding creates one or more 3D objects containing the highest cumulative density, and the voxel v c Each of the vertices corresponds to the center of such an object. This strategy is advantageous because most of the gradients in the cochlea do not converge to one voxel, but to several neighboring voxels. Therefore, to improve the accuracy of the detection (especially in noisy images), it is preferable to search for the object with the highest intensity.

[0018] According to one embodiment of the method, forming the cochlear descriptor image includes calculating a Hadamard product (E1 ○ E2 ○ E3) ij =(E1) ij (E2) ij (E3) ij Calculate.

[0019] Further, according to one embodiment of the method, each gradient line has a length extending from a respective starting point to a respective end point, the length of each first gradient line being greater than a length of each third gradient line, and the length of each third gradient line being greater than a length of each second gradient line.

[0020] Further, according to one embodiment of the method, each of the gradient lines includes a length extending from the respective starting point to the respective end point, the length of each of the first gradient lines being greater than the length of each of the third gradient lines, and the length of each of the third gradient lines being greater than the length of each of the second gradient lines.

[0021] Further, according to one embodiment of the method, each of the first, second and third starting points p start,1 , p start,2 , p start,3 is calculated according to p start,i =position(v M )+start i gi=1, 2, 3 Each of the first, second and third end points p end,1 , p end,2 , p end,3 , is calculated according to p end,i =position(v M )+end i gi=1, 2, 3 Here, position(v M ) indicates the location of each voxel (v) in the image (M). Furthermore, according to one embodiment of the method, start1 is in the range of 2.7 mm to 3.3 mm, in particular start1 is equal to 3.0 mm; Furthermore, according to one embodiment of the method, start2 ranges from −3.3 mm to −2.7 mm, and in particular start1 is equal to −3.0 mm. Furthermore, according to an embodiment of the method, start3 is in the range of 0.45 mm to 0.55 mm, in particular start3 is equal to 0.5 mm. Furthermore, according to one embodiment of the method, end1 is in the range of 4.5 mm to 5.5 mm, in particular end1 is equal to 5.0 mm; Furthermore, according to an embodiment of the method, end2 is in the range of −1.8 mm to −2.2 mm, in particular end2 is equal to −2.0 mm; Furthermore, according to an embodiment of the method, end3 is in the range of 1.8 mm to 2.2 mm, in particular end3 is equal to 2.0 mm.

[0022] Moreover, in one embodiment, the constant increment corresponds to the natural number one.

[0023] Moreover, according to one embodiment of the method, each gradient vector is a normalized gradient vector.

[0024] Further, according to an embodiment of the method, the image (M) is one of a CT image, an MRI image, and a cone beam CT image.

[0025] Moreover, according to one embodiment of the method, the method further comprises the steps of: - providing a statistical geometric model of the human ear including at least a cochlear portion corresponding to the cochlea of ​​the human ear; In particular, according to one embodiment of the present invention, the SSM includes at least one, some, or all of the following additional parts of the ear in addition to the cochlea: Three semicircular canals, stapes, incus, malleus, and tympanic membrane. Statistical Shape Models (SSM) are known in image analysis. In the context of the present invention, SSM are used in particular for segmenting images of the cochlea. The SSM comprises / describes the average shape of at least the cochlea (and in particular other parts present in the SSM) and shows the variation of the training data sets with respect to the average shape. In particular, according to one embodiment of the invention, the average shape is calculated from at least 20 training data sets, each training data set corresponding to a 3D image of a different human ear, the 3D image comprising at least the cochlea part and in particular the other part(s) (see above). In particular, to compute the variance, rotation and translation are removed from each training data set. More specifically, Principal Component Analysis (PCA) is used to extract variation from the training dataset by computing a covariance matrix from the training dataset matrix that includes the training dataset. The eigenvectors corresponding to the eigenvalues ​​of the covariance matrix correspond to the direction and amount of variation found in the training dataset. The principal modes of variation are calculated based on the first N eigenvectors and their associated eigenvalues, where the number N depends on the desired cumulative variance, i.e., the value of the variation that can be explained by the first N eigenvectors. The possible shapes x of the SSM are then calculated as a weighted linear combination of these eigenvectors Ψ around the mean shape x, x = x - +Ψ w and the corresponding weights w are denoted as feature weights. (x - is the mean of x)

[0026] According to yet another embodiment, the method further comprises the following steps: For each voxel (v c ) ○For each voxel (v c ) defining a volume in said image (M) centered on o Segment the cochlea in the volume using a Hessian-based enhancement filter, thereby obtaining a segmented cochlear region of said image M, said region corresponding to the cochlea in image M, in particular the Hessian-based filter can be configured to detect dark tubular structures, in particular α=0.1, β=0.5, and γ=100. The filter is configured in a multi-scale mode, for example with σ varying between 0.3 mm and 0.7 mm in three steps, α=0.3, β=0.5, γ=0.7. In particular, the segmented cochlear region can be converted into a 3D mesh, for example by applying a marching cubes algorithm (e.g. William E. Lorensen, Harvey E. Cline: Marching Cubes: A high resolution 3D surface construction algorithm. In: Computer Graphics, Vol. 21, No. 4, July 1987). This allows obtaining a 3D surface of the cochlea. o Fitting the cochlear portion of a statistical shape model to the 3D surface of the cochlea. In particular, the fitting can involve testing a range of cochlear sizes, i.e., varying the initial variation mode that determines the size (scale), in particular for x = x - +Ψ w The following values ​​of the first entry of w in are tested: −3.0, −2.0, −1.0, 0.0, 1.0, 2.0, 3.0. Now, for every generated cochlea, a rotation is applied around its central axis to find the optimal orientation. Calculating a surface similarity metric (e.g. root mean square difference) indicative of the similarity between the surface of said cochlear portion of the statistical shape model and the surface of the segmented cochlear region of the image (M), - voxel v of said number P c Among voxels v, the surface similarity index satisfies a predefined criterion. c as the center of the cochlea, and in particular, the first voxel v among the number P of voxels whose surface similarity index is below a predefined threshold. cis selected as the center of the cochlea. In particular, image segmentation corresponds to classifying whether a voxel of an image belongs to a particular structure or not. In particular, the following Hessian-based enhancement filter may be used in this context: Furthermore, each volume may be a rectangular or spherical volume, but may also consist of any other suitable shape, in particular the volumes are cuboids with side lengths of at least 20 mm.

[0027] Moreover, according to one embodiment of the method, the method further comprises the steps of: - fitting the cochlea portion of the statistical shape model to the segmented cochlea region using the determined center of the cochlea, the statistical shape model further including semicircular canal portions corresponding to the three semicircular canals of the ear. - predicting positions of the three semicircular canals in the image M as the positions of the three semicircular canals in the semicircular canal portion of the statistical shape model; - defining a volume that contains the predicted location; - segmenting said three semicircular canals in said volume (for example using a Hessian-based enhancement filter), thereby obtaining segmented semicircular canal regions of image M that correspond specifically to the three semicircular canals in image M. In particular, the Hessian-based enhancement filter used is configured to detect dark tubular structures, in particular α=0.5, β=0.1, γ=100, with σ ranging from 0.3 mm to 0.7 mm in three steps (for example σ=0.3, σ=0.5, σ=0.7).

[0028] Furthermore, based on the statistical shape model and fitting to the segmented cochlear region and the segmented semicircular canal region, the three semicircular canals in image M may be annotated. Again, the volumes used may be rectangular volumes, in particular cubes, spherical volumes or any other suitable volumes.

[0029] Moreover, according to one embodiment of the method, the method further comprises the steps of: - fitting the cochlea and semicircular canal portions of the statistical shape model to the segmented cochlea region corresponding to the cochlea in the image and to the segmented semicircular canal regions corresponding in particular to three semicircular canals in the image M. Advantageously, this step improves the fit of the ear SSM shape to the anatomical structures, making further predictions more accurate.

[0030] Moreover, according to one embodiment of the method, the method further comprises the steps of: - predicting positions of the incus, malleus and stapes in the image as positions of the incus, malleus and stapes in an ossicular region of the statistical shape model, the ossicular region of the statistical shape model corresponding to the incus, malleus and stapes of the ear; - defining a volume that contains the predicted location; - segmenting the incus, malleus and stapes in said volume, thereby obtaining segmented ossicular regions of said image M which correspond to the incus, malleus and stapes in said image M. Furthermore, in accordance with an embodiment of the present invention, the incus, malleus, and stapes may be annotated in the image based on the statistical shape model and fitting to the segmented cochlea region, the segmented semicircular canal region, and the segmented ossicular region. Again, the volumes used may be rectangular volumes, in particular cubes, spherical volumes or any other suitable volumes.

[0031] Moreover, according to one embodiment of the method, the method further comprises the steps of: - predicting a location of the tympanic membrane as the location of the tympanic membrane in a tympanic membrane portion of the statistical shape model, the tympanic membrane portion of the statistical shape model corresponding to the tympanic membrane of the ear; - defining a volume that contains the predicted location; - segmenting the tympanic membrane within the volume (e.g., using a Hessian-based enhancement filter), thereby obtaining a segmented tympanic membrane region in the image M that corresponds to the tympanic membrane in the image M; In particular, this Hessian filter may be configured to detect bright plate-like structures, in particular with α=0.1, β=1.0, γ=50, σ=0.2-0.4, with two stages, for example σ=0.2 and σ=0.4. Further, according to one embodiment, the tympanic membrane may be annotated in the image based on a statistical shape model and fitting to at least the segmented cochlea region, the segmented semicircular canal region, and the segmented tympanic membrane region. Again, the volumes used may be rectangular volumes, in particular cubes, spherical volumes or any other suitable volumes. In particular, the tympanic membrane is the terminus of the ear canal. Once the tympanic membrane is segmented, we know where the ear canal ends and its orientation (in particular, a best-fit plane for the tympanic membrane is sufficient to predict the orientation of the ear canal). In particular, in one embodiment, once the ear canal termination and its orientation are known, a cylindrical region of interest (the shape that best represents the ear canal) is defined. Then, in the defined shape, the cortical bone representing the ear canal is segmented. This bone can be segmented using a simple thresholding or a Hessian filter (set to light plate-like structures).

[0032] Moreover, according to one embodiment of the method, the method further comprises the steps of: - predicting a region, a direction and a boundary of the location of the ear canal based on the segmented tympanic membrane region in image M corresponding to the tympanic membrane in image M; - defining a volume that contains the predicted region; - segmenting the ear canal in the volume (e.g. using a Hessian-based enhancement filter), thereby obtaining a segmented ear canal region of the image M that corresponds to the ear canal in the image M. Further, in one embodiment, the ear canal may be annotated in the image based on a statistical shape model and fitting to at least the segmented tympanic membrane region. Again, the volumes used may be rectangular volumes, in particular cubes, spherical volumes or any other suitable volumes.

[0033] Moreover, according to one embodiment of the method, the method further comprises the steps of: - predicting a position and an orientation of said outer surface of the temporal bone based on a segmented ear canal region of image M corresponding to the ear canal in image M; - defining a volume that contains the predicted location; - segmenting said temporal bone in said volume using a thresholding process, thereby obtaining a segmented temporal bone region of the image (M) consisting of the outer surface of the temporal bone (in particular the adjacent layer with a thickness of about 5 mm) and detecting said outer surface using said orientation of said ear canal, in particular the surface with a normal pointing in the direction of the ear canal (within a deviation of 40 degrees or less) being considered as the outer surface; Furthermore, according to one embodiment, annotation of the temporal bones in the images may be performed. Again, the volumes used may be rectangular volumes, in particular cubes, spherical volumes, or any other suitable volumes.

[0034] Moreover, according to one embodiment of the method, the method further comprises the steps of: - predicting a location of the inner ear canal based on a segmented cochlear region of the image corresponding to the cochlea in the image; - defining a volume that contains the predicted location; - segmenting an ear canal in said volume, thereby obtaining a segmented ear canal region of said image M corresponding to the ear canal in said image; Furthermore, in one embodiment of the present invention, the inner ear canal may be annotated in the image. Again, the volume used may be a rectangular volume, in particular a cube, a spherical volume, or any other suitable volume.

[0035] Moreover, according to one embodiment of the method, the method further comprises the steps of: - predicting a location of the facial nerve based on the segmented cochlea region of the image M corresponding to the cochlea and the segmented semicircular canal regions of the image M corresponding to three semicircular canals; - defining a volume that contains the predicted location; - Segment the facial nerve in said volume (eg using a Hessian-based enhancement filter), thereby obtaining a segmented facial nerve region in image M that corresponds to the facial nerve in said image. Furthermore, according to one embodiment, annotation of the facial nerve in the segmented images is performed.

[0036] Further, according to one embodiment of the method, the method includes graphically visualizing for a user on a display (102) at least one of the segmented cochlea region, the segmented semicircular canal region, the segmented ossicular region, the segmented tympanic membrane region, the segmented external ear canal region, the segmented temporal bone region, the segmented internal ear canal region, and the segmented facial nerve region.

[0037] Furthermore, according to a further aspect of the present invention, a system for automatically segmenting a human cochlea contained in an input image M composed of voxels, each voxel including an intensity value indicative of an intensity of the respective voxel, is disclosed, the system comprising at least one processor and a display, the system being adapted to perform the method according to the present invention. In particular, the system is configured to display one, some or all of the following segmented regions of image M on the display: a segmented cochlea region, a segmented semicircular canal region, a segmented ossicular region, a segmented tympanic membrane region, a segmented external ear canal region, a segmented temporal bone region, a segmented internal ear canal region, a segmented facial nerve region. In particular, each segmented region may be overlaid on the image M on the display and / or displayed separately.

[0038] According to a further aspect of the invention, a computer program is disclosed, which comprises instructions for causing a system according to the invention to carry out the steps of the method according to the invention when the program is executed on the system according to the invention.

[0039] According to a further aspect of the invention, there is provided a computer-readable, in particular non-transitory, storage medium comprising instructions which, when executed by a system (or a computer) according to the invention, cause the system (or computer) according to the invention to perform the steps of the method according to the invention.

[0040] Furthermore, according to yet another aspect, the invention relates to a data carrier signal carrying a computer program according to the invention. [Brief description of the drawings]

[0041] Further features and advantages of the present invention and embodiments of the present invention are described below with reference to the drawings. [Figure 1] Figure 1 shows a diagram of the cochlea and its characteristic shape with different circles. [Diagram 2] FIG. 2 illustrates the geometric principle of automatically detecting the center of the cochlea. [Diagram 3] FIG. 3 shows a schematic representation of an empty first gradient accumulation image E1 used to detect the center of the cochlea in a medical image. [Figure 4]FIG. 4 shows how the ear SSM is fitted to the cochlea in the image based on the detected cochlea center using a surface similarity metric that can also detect whether the cochlea belongs to the left or right ear. [Diagram 5] FIG. 5 shows the prediction of the semicircular canal location and its segmentation based on the segmented cochlea. [Figure 6] FIG. 6 shows labeling of the semicircular canals based on the ear SSM adapted to the segmented cochlea and semicircular canals. [Figure 7] FIG. 7 shows the prediction of the tympanic membrane location and the segmentation of the tympanic membrane. [Figure 8A] FIG. 8A, for example, shows a projection of the temporal bone. [Figure 8B] FIG. 8B, for example, shows the facial nerve projection. [Figure 9] FIG. 9 illustrates one embodiment of a system according to the present invention. [Figure 10] FIG. 10 illustrates one embodiment of a system according to the present invention.

[0042] The present invention provides a method for automatically detecting and segmenting the human cochlea in an image M composed of voxels v containing intensity values ​​indicative of the intensity of each voxel v. Such an image may be acquired by suitable medical imaging equipment, which may be part of a system according to the present invention.

[0043] To find the center 11 of the cochlea 10, the voxel v E1 An empty first gradient accumulation image E1 consisting of E1 is associated with a voxel v of image M. Then, for each voxel v of image M, a gradient vector g is calculated that points in the direction of the largest possible intensity increase in image M, and a first gradient line is calculated that extends in the direction of the gradient vector g, the first gradient line starting from a first starting point p start,1 Starts at the first end point p end,1(see FIG. 2 ). Here, for each voxel v in the first gradient accumulation image E1, where the corresponding voxel v in image M intersects with the respective first gradient line, E1 The intensity value of is increased by a predefined constant increment.

[0044] Furthermore, voxel v E2 (see FIG. 3) is generated, and each voxel v of the second gradient accumulation image E2 is E2 is also associated with voxels v of image M, and for each voxel v of image M, a gradient vector g is calculated that points in the direction of the largest possible intensity increase in image M, and a second gradient line is calculated that runs in the opposite direction to the gradient vector g, the second gradient line starting from a second starting point p start,2 Starts at the second end point p end,2 (see FIG. 2 ) and for each voxel v in the second gradient accumulation image E2 where the corresponding voxel v in image M intersects the respective second gradient line. E2 The intensity value of is increased by a predefined constant increment.

[0045] Similarly, for voxel v E3 An empty third gradient accumulation image E3 is provided (see FIG. 3), where each voxel v E3 are in turn associated with voxels v of image M, and for each voxel v of image M, a gradient vector g pointing in the direction of the maximum possible intensity increase in image M is calculated, and a third gradient line extending in the direction of gradient vector g is calculated, the third gradient line starting from a third starting point p start,3 Starts at the third end point p end,3 (see FIG. 2 ), and for each voxel v in the first gradient accumulation image E3 where the corresponding voxel v in image M intersects the respective third gradient line. E3 The intensity value of is increased by a predefined constant increment.

[0046] Next, the first, second and third gradient accumulation images E1, E2, E3 are composited to obtain voxel v c A cochlear descriptor image E consisting of descriptor and the intensity values ​​are stored as cochlear descriptor images Edescriptor voxel v with P ≥ 1 that is greater than the intensity value of other voxels in c is selected as a predictor of the location of the center of the cochlea in said image M. In particular, the above procedure can be employed to determine candidates for the center 11 of the cochlea 10.

[0047] The reason for using three gradient accumulation images E1, E2, E3 is based on the observation that in general the shape of the cochlea can be described as a snail shape, consisting of two and a half revolutions, and therefore a relatively large number of normals of the cochlear surface points pass through the actual center 11 of the cochlea 10.

[0048] Therefore, the main principle behind the cochlear center detection algorithm can be understood as the task of finding the point in space that contains the maximum number of surface normal intersections.

[0049] This principle is illustrated in Figure 2 for finding the center of a circle with a known radius, where a short line is defined that starts before the center of the circle and ends after the center of the circle. Such an algorithm allows to find circles of a defined size.

[0050] This principle can be exploited to design an algorithm for cochlear center detection since the mean radius of the cochlea and its variation are known.

[0051] Below is the pseudocode for finding the center of a circle as an example. The algorithm requires the following as input: - an image M containing the detected center (e.g. the center of the cochlea, here the center of the circle for simplicity), this image M can be a CT, MR or cone beam CT volumetric image containing the circle / cochlea. - An empty image E, specifically having the same origin and size as M. The image spacing (voxel size) may differ from image M. This image is called the cumulative image. - Start distance - The distance from the voxel position to the start of the line (along the voxel's gradient direction). - End distance- The distance from the voxel position to the end point of the line (along the voxel's gradient direction). Then, an accumulated gradient vector intersection image is generated as an output. The corresponding pseudocode can be formulated as follows: JPEG2025507879000002.jpg84107

[0052] The circle center detection algorithm can be easily adapted to the cochlear situation first described by understanding that the shape of the cochlea defines three circles, shown diagrammatically in Figure 1.

[0053] In particular, the cochlea can be modeled as three distinct circles: The first circle C1 captures the outer wall of the cochlea's base 10. Anatomically, it represents the transition from the cochlear bone (high intensity) to the cochlear fluid (low intensity), with these gradient vectors pointing towards the cochlea centre 11. The second circle C2 captures the inner wall of the cochlea's base 10. Anatomically, it represents the transition from cochlear fluid (low intensity) to cochlear bone (high intensity), and these gradient vectors point outward from the cochlear center 11, but are centrally aligned. The third circle C3 captures the outer wall of the second and second-and-a-half turns of the cochlea. Anatomically, it represents the transition from the cochlear bone (high intensity) to the cochlear fluid (low intensity), with these gradient vectors again pointing towards the cochlear centre 11.

[0054] In pseudocode, this can be implemented as follows: Given an input image M, the cumulative image E for each circle i Calculate. That is, the accumulation value image of the circle C1 is as follows. E1= Circle Center Detection Algorithm (M, 3.0, 5.0). For the second circle C2, the gradient direction is outward so a negative distance is used. E2= Circle Center Detection Algorithm (M, -3.0, -2.0). For the third circle C3 we have: E3= Circle Center Detection Algorithm (M, 0.5, 2.0). The cochlear descriptor image is calculated, for example, as a multiplication of three accumulated images. E descriptor =E1*E2*E3 Here, for example, E descriptor The brightest voxel of may be selected as a representative of the center 11 of the cochlea 10. As described herein, other procedures for selecting candidates for the cochlea center 11 may also be applied.

[0055] A preferred variant of finding a candidate for the center 11 of the cochlea 10 is to multiply the brightest voxel of the cochlea descriptor image E by a factor ranging from 0.6 to 0.8, preferably 0.7. descriptor The key to success is to threshold the cochlear center 11, i.e. only the brightest parts of the corresponding voxels are considered. This typically results in a large number of 3D objects in the descriptor image that constitute the highest cumulative density, and each candidate for the cochlear center 11 is the center of such an object.

[0056] As shown in FIG. 4, in order to find the true center 11 of the cochlea 10, in one embodiment of the present invention, these possible centers of the cochlea can be tested as follows. a. Crop a box B1 around each candidate cochlear center, for example with dimensions 20x20x20 mm (other volume shapes and sizes may be used). b. In Box B1, the tubular structure is automatically segmented using a Hessian-based enhancement filter, and preferably the segmented structure is converted into a 3D mesh by applying a marching cubes algorithm to obtain a 3D surface. c. Automatically fitting a statistical shape model (SSM) of the ear, i.e. only the cochlear part, to the segmented data D1, in particular by fitting the cochlear part to said 3D surface, left and right cochlear SSMs are fitted, since it is initially unknown which side (left or right ear) the cochlea belongs to. d. Based on the surface similarity metric, it is automatically determined which candidate represents the center 11 of the cochlea 10 (i.e., has the best surface similarity), which also gives the side of the cochlea (left or right).

[0057] Furthermore, FIG. 5 shows a prediction of the position of the semicircular canals 12 and their segmentation based on the segmented cochlea 10. a. To automatically predict and segment the location of the semicircular canals 12, the ear SSM is automatically adapted to the segmented cochlea 10 to obtain a prediction of the location of the semicircular canals 12. b. Next, a volume such as box B2 is cropped around the prediction region. c. Finally, segmentation is automatically performed within the volume to segment the semicircular canals 12 within the image M.

[0058] As shown in FIG. 6, by fitting an ear SSM to the cochlea 10 and semicircular canals 12, the semicircular canals 12 can be labeled (annotated).

[0059] Furthermore, the cochlea and semicircular canal portions of the statistical shape model SSM can be automatically fitted to the segmented cochlea 10 and the segmented semicircular canals 12, based on which the positions of the incus 5, malleus 6 and stapes 7 (see also FIG. 8A) can be predicted within volume B4 and then segmented within said volume B4.

[0060] Furthermore, once the ear SSM has been adapted to the cochlea 10 and semicircular canals 12, the approximate location of the tympanic membrane 13 is also known. This is exemplarily shown in Fig. 7. Again, a volume B3 is defined containing the predicted location, and the tympanic membrane 13 is automatically segmented within volume B3, thereby resulting in a segmented tympanic membrane region in image M.

[0061] Furthermore, FIG. 8A shows the inner ear 1 and middle ear 2 together with the ear canal 3 and the pinna 4. As shown in FIG. 8A, the area, orientation and boundary of the location of the ear canal 14 can be predicted based on the shape of the segmented eardrum (see above). After defining a volume B5 containing the predicted area, the ear canal 14 can be automatically segmented in the volume B5. Based on the segmented ear canal 14, the location and orientation of the outer surface 15 of the temporal bone 16 can be automatically predicted. After defining a volume B6 containing the predicted location, the temporal bone 16 can be automatically segmented in the volume B6 using a thresholding process (bones are always strong). The orientation of the ear canal 14 can be used to detect the outer surface 15.

[0062] Furthermore, because the inner ear canal 17 is located below the cochlea 10, the location of the inner ear canal 17 can also be automatically predicted based on the segmented cochlea 10. By defining a volume B7 that contains the predicted location, the inner ear canal can be automatically segmented within volume B7.

[0063] Furthermore, the location of the facial nerve 18 can be automatically predicted based on the segmented cochlea 10 and the segmented semicircular canals 12. As shown in FIG 8B, a volume B8 is defined that contains the predicted location, and the facial nerve 18 can be automatically segmented within volume B8.

[0064] The segmented images may be advantageously used for a variety of different applications, such as surgical planning, intraoperative monitoring, image-guided surgery, diagnosis and patient education, particularly for presenting the planned surgery and / or treatment to a patient pre-operatively.

[0065] 9 shows an embodiment of a system 100 according to the invention configured to automatically detect and segment the cochlea 10 of a human ear in an input image M composed of voxels v, each voxel comprising an intensity value indicative of the intensity of the respective voxel. Preferably, the system 100 comprises at least one processor 101 and a display 102, the at least one processor 101 and the display 102 being adapted to perform the method according to the invention.

[0066] Preferably, the system is configured to display one, some or all of the following segmented regions of the image M on the display 102: a segmented cochlea region 10 and / or its centre 11, a segmented semicircular canal region 12, segmented ossicular regions 5, 6, 7, a segmented tympanic membrane region 13, a segmented external ear canal region 14, a segmented temporal bone region 16, a segmented internal ear canal region 17, a segmented facial nerve region 18.

[0067] The system 100 is preferably configured to allow input of input data such as the image(s) M and the statistical shape model SSM. The system 100 according to the invention can be realized by a general purpose computer and a display connected thereto, on which computer a computer program can be executed to cause the computer to execute the method according to the invention. However, the system can also be composed or configured with hardwired components / functions configured to execute the method according to the invention. Furthermore, FIG. 10 shows an embodiment of the system 100 according to the invention, adapted to perform an image-guided surgery. Apart from the components described above in relation to FIG. 9, the system 100 further comprises at least one surgical instrument 103 for performing surgery on the patient's ear, in particular on one of the segmented parts of the ear, such as the cochlea 10, the semicircular canals 12, the ossicles 5, 6, 7, the tympanic membrane 13, the external ear canal 14, the temporal bone 16, the internal ear canal 17, the facial nerve 18. In particular, the surgical instrument is driven by an actuator configured to be controlled by the surgeon or driven fully automatically via a manipulation device. The manipulation devices may be connected to at least one processor / computer 101 via an interface, and the actuators may be controlled via the at least one processor / computer 101 .

[0068] In particular, herein, system 100 is preferably configured to perform one or some or all of the following functions: - automatically segmenting and specifically annotating, in at least one image M, at least one structure or structures of the ear as described herein (such as the cochlea 10, the semicircular canals 12, the ossicles 5, 6, 7, the tympanic membrane 13, the external auditory canal 14, the temporal bone 16, the internal auditory canal 17, the facial nerve 18, etc.) using a statistical shape model of the ear; - automatically rendering (e.g., on the display 102) a graphical visualization of said at least one structure or structures; - automatically registering said at least one structure or said plurality of structures to a current location of a patient; - display of such registered structures, especially in combination with a real-time video representation of the patient during the operation (e.g. on the display 102 and / or on a further display) - allowing automatic tracking of at least one surgical instrument relative to the patient and to the registered structure(s), in particular a perfect visualization of the surgical site.

[0069] Further aspects of the present invention and embodiments thereof are listed below as items which may also be presented as claims of the present invention. Reference numerals in parentheses refer to the figures described herein.

[0070] Item 1: A method for automatically segmenting the human cochlea (10) in an image (M) comprised of voxels (v), each voxel including an intensity value indicative of the intensity of the respective voxel (v), the method comprising: - Voxel(v E1 ), wherein each voxel (v E1 ) provides the first gradient accumulation image (E1) associated with voxels (v) of the image (M), wherein for each voxel (v) of the image (M), a gradient vector (g) pointing in the direction of the largest possible intensity increase of the image (M) and further a first gradient line extending in the direction of the gradient vector (g), the first gradient line starting from a first starting point p start,1 Starts at the first end point p end,1 and for each voxel (v) of the first gradient accumulation image (E1) where a corresponding voxel (v) of the image (M) intersects with each of the first gradient lines, the first gradient lines ending with E1 ) by a predefined constant increment; - Voxel(v E2 ), wherein for each voxel (v E2) provides the second gradient accumulation image (E2) associated with voxels (v) of the image (M), comprising for each voxel (v) of the image (M) a gradient vector (g) pointing in the direction of the greatest possible intensity increase of the image (M) and further a second gradient line extending in the opposite direction of the gradient vector (g), the second gradient line starting from a second starting point p start,2 Starts at the second end point p end,2 and for each voxel (v) of the second gradient accumulation image (E2) where a corresponding voxel (v) of the image (M) intersects with the respective second gradient line, the second gradient line ends at E2 ) by a predefined constant increment; - Voxel(v E3 ), wherein for each voxel (v E3 ) provides the third gradient accumulation image (E3) associated with voxels (v) of the image (M), wherein for each voxel (v) of the image (M), a gradient vector (g) pointing in the direction of the greatest possible intensity increase of the image (M) and further a third gradient line extending in the direction of the gradient vector (g), the third gradient line starting from a third starting point p start,3 Starts at the third end point p end,3 and for each voxel (v) of the first gradient accumulation image (E3) where a corresponding voxel (v) of the image (M) intersects with each of the third gradient lines, the third gradient lines ending with E3 ) by a predefined constant increment; Using the first, second and third gradient accumulation images (E1, E2, E3), a voxel (v c ) is composed of cochlear descriptor images (E descriptor ), wherein the intensity values ​​correspond to the cochlear descriptor image (E descriptor ) is greater than the intensity values ​​of all other voxels of the cochlear descriptor image (E descriptor ) number of voxels (v c) and using it to predict the location of the center (11) of the cochlea (10) in said image (M); A method comprising:

[0071] Item 2: The cochlear descriptor image (E descriptor ) is a Hadamard product (E1 ○ E2 ○ E3) ij =(E1) ij (E2) ij (E3) ij 2. The method according to claim 1, which corresponds to calculating

[0072] Item 3: Each of the gradient lines is start,1 , p start,2 , p start,3 ) to each of the end points (p end,1 , p end,2 , p end,3 ), the length of each first gradient line being greater than the length of each third gradient line, and the length of each third gradient line being greater than the length of each second gradient line.

[0073] Item 4: The first, second and third starting points p start,1 , p start,2 , p start,3 is calculated according to p start,i =position(v)+start i gi=1, 2, 3 Each of the first, second and third end points p end,1 , p end,2 , p end,3 , is calculated according to p end,i =position(v)+end i gi=1, 2, 3 where position(v) denotes the position of the respective voxel (v) in the image (M); - start1 is in the range from 2.7mm to 3.3mm, specifically start1 is equal to 3.0mm, - start2 ranges from -3.3mm to -2.7mm, and in particular start1 is equal to -3.0mm, - start3 is in the range from 0.45mm to 0.55mm, specifically, start3 is equal to 0.5mm, - end1 is in the range from 4.5mm to 5.5mm, specifically end1 is equal to 5.0mm, - end2 is in the range of -1.8mm to -2.2mm, specifically end2 is equal to -2.0mm, - end3 is in the range of 1.8mm to 2.2mm, specifically end3 is equal to 2.0mm, 2. The method according to any of the preceding items.

[0074] Item 5: A method according to any one of the preceding items, wherein the image (M) is one of a CT image, an MRI image, and a cone beam CT image.

[0075] Item 6: providing a statistical geometric model (SSM) of an ear including at least a cochlear portion, said cochlear portion corresponding to said cochlea (10) of said ear; The method of any one of the preceding claims, further comprising:

[0076] Item 7: For each voxel (v c ) Each of the voxels (v c ) defining a volume (B1) in said image (M) centered on segmenting the cochlea (10) in said volume (B1) using a Hessian-based enhancement filter, thereby obtaining a segmented cochlear region of said image (M), and generating a 3D surface of the cochlea from said segmented cochlear region, in particular by using a marching cubes algorithm; fitting the cochlear portion of the statistical shape model (SSM) to the 3D surface; calculating a surface similarity metric indicating the similarity between the surface of the cochlear portion of the statistical shape model (SSM) and the 3D surface, - the number P of voxels (v c ), the surface similarity index satisfies a predefined criterion. c ) as the center (11) of the cochlea (10), and in particular the first voxel (v c ) is selected as the center of the cochlea, The method according to item 6, further comprising:

[0077] Item 8: - fitting the cochlea portion of the statistical shape model (SMM) to the segmented cochlea region using the determined center (11) of the cochlea (10), the statistical shape model (SSM) further comprising semicircular canal portions corresponding to the three semicircular canals (12) of the ear; - predicting the positions of the three semicircular canals (12) in the image (M) as the positions of the three semicircular canals (12) in the semicircular canal portion of the statistical shape model (SSM); - defining a volume (B2) that contains said predicted location, - segmenting said three semicircular canals (12) within said volume (B2), thereby obtaining segmented semicircular canal regions of said image (M); 8. The method of claim 7, further comprising:

[0078] Item 9: - fitting the cochlea and semicircular canal portions of the statistical shape model (SSM) to the segmented cochlea and semicircular canal regions; The method of claim 8, further comprising:

[0079] Item 10: - predicting the positions of the incus (5), malleus (6) and stapes (7) in the image (M) as the positions of the incus (5), malleus (6) and stapes (7) in the ossicular region of the statistical shape model (SSM), the ossicular region of the statistical shape model corresponding to the incus (5), malleus (6) and stapes (7) of the ear, - defining a volume (B4) that contains said predicted location, - segmenting the incus, malleus and stapes in the volume, thereby obtaining segmented ossicular regions of the image (M) corresponding to the incus, malleus and stapes in the image (M); 10. The method of claim 9, further comprising:

[0080] Item 11: - predicting a position of the tympanic membrane (13) as the position of the tympanic membrane (13) in a tympanic membrane portion of the statistical shape model (SSM), the tympanic membrane portion of the statistical shape model corresponding to the tympanic membrane (13) of the ear; - defining a volume (B3) that contains said predicted location, - segmenting said tympanic membrane (13) within said volume (B3), thereby obtaining a segmented tympanic membrane region of said image (M); 11. The method of claim 9 or 10, further comprising:

[0081] Item 12: - predicting the area, orientation and boundary of the location of the ear canal (14) based on the segmented eardrum region of the image (M); - defining a volume (B5) containing said predicted area, - segmenting said ear canal (14) within said volume (B5), thereby obtaining a segmented ear canal region of said image (M); 12. The method of claim 11, further comprising:

[0082] Item 13: - predicting the position and orientation of the outer surface (15) of the temporal bone (16) based on the segmented ear canal region of the image (M); - defining a volume (B6) that contains said predicted location, - segmenting the temporal bone (16) in the volume (B6) using a thresholding process, thereby obtaining a segmented temporal bone region in the image (M) and detecting the outer surface using the orientation of the ear canal (14), 13. The method of claim 12, further comprising:

[0083] Item 14: - predicting a location of the inner ear canal (17) based on the segmented cochlear region in the image; - defining a volume (B7) that contains said predicted location, - segmenting said ear canal (17) within said volume (B7), thereby obtaining a segmented ear canal region of said image (M); The method according to item 7, or the method according to any one of items 8 to 13 based on item 7, further comprising:

[0084] Item 15: - predicting a location of the facial nerve (18) based on the segmented cochlea region, the segmented semicircular canal region, and the segmented ear canal region of the image (M); - defining a volume (B8) that contains said predicted location, - segmenting said facial nerve (18) within said volume (B8), thereby obtaining a segmented facial nerve region of said image (M); The method according to items 7, 9 and 12, further comprising:

[0085] Item 16: 16. The method of any one of items 7 to 15, further comprising graphically visualizing for a user on a display (102) at least one of the segmented cochlea region, the segmented semicircular canal region, the segmented ossicular region, the segmented tympanic membrane region, the segmented external ear canal region, the segmented temporal bone region, the segmented internal ear canal region, and the segmented facial nerve region.

[0086] Item 17: A system (100) for automatically segmenting a human cochlea (10) contained in an input image (M) composed of voxels (v), each voxel including an intensity value indicative of the intensity of the respective voxel, the system (100) comprising at least one processor (101) and a display (102), the at least one processor (101) and the display (102) being adapted to perform the method according to any one of items 1 to 16.

[0087] Item 18: A computer program comprising instructions that, when executed on a system (100) according to item 17, cause the system (100) to perform the steps of the method according to one of items 1 to 16.

[0088] Item 19: A computer-readable storage medium comprising instructions that, when executed by the system according to claim 17, cause the system according to claim 17 to perform the steps of the method according to any one of claims 1 to 16.

[0089] Item 20: 19. A data carrier signal carrying a computer program according to item 18.

Claims

1. A method for automatically segmenting the human cochlea (10) contained in an image (M) composed of voxels (v), each voxel including an intensity value indicative of the intensity of the respective voxel (v), the method comprising: - Voxel (v Ｅ１ ) is constructed from the first empty gradient accumulation image (E １ ), wherein the first gradient accumulation image (E １ ) for each voxel (v Ｅ１ ) is associated with a voxel (v) of said image (M), said first gradient accumulation image (E １ ), wherein for each voxel (v) of said image (M), a gradient vector (g) pointing in the direction of the greatest possible intensity increase of said image (M), and further a first gradient line extending in the direction of said gradient vector (g), said first gradient line starting from a first starting point p start,１ Starting at the first end point p ｅｎｄ，１ and computing the first gradient lines ending with , and computing the first gradient accumulation images (E １ ) for each voxel (v Ｅ１ ) by a predefined constant increment; - Voxel (v Ｅ２ ) is constructed from the empty second gradient accumulation image (E ２ ), wherein the second gradient accumulation image (E ２ ) for each voxel (v Ｅ２ ) is the second gradient accumulation image (E ２ ), wherein for each voxel (v) of said image (M), a gradient vector (g) pointing in the direction of the greatest possible intensity increase of said image (M) and further a second gradient line extending in the opposite direction of said gradient vector (g), said second gradient line starting from a second starting point p start,２ Starting at the second end point p ｅｎｄ，２ and calculating the second gradient lines ending with , and calculating the second gradient accumulation images (E ２ ) for each voxel (V Ｅ２ ) by a predefined constant increment; - Voxel (v Ｅ３ ) and an empty third gradient accumulation image (E ３ ), wherein the third gradient accumulation image (E ３ ) for each voxel (v Ｅ３ ) is associated with a voxel (v) of said image (M); ３ ), wherein for each voxel (v) of said image (M), a gradient vector (g) pointing in the direction of the greatest possible intensity increase of said image (M), and further a third gradient line extending in the direction of said gradient vector (g), said third gradient line starting from a third starting point p start,３ Starting at the third end point p ｅｎｄ，３ and computing the third gradient lines, each ending with a corresponding voxel (v) of the image (M) at which the first gradient accumulation image (E ３ ) for each voxel (v Ｅ３ ) by a predefined constant increment; the first, the second and the third gradient accumulation images (E １ , E ２ , E ３ ) to obtain the voxel (v ｃ ) is composed of a cochlear descriptor image (E ｄｅｓｃｒｉｐｔｏｒ ), wherein the intensity values ​​are included in the cochlear descriptor image (E ｄｅｓｃｒｉｐｔｏｒ ) is greater than the intensity values ​​of all other voxels of the cochlear descriptor image (E ｄｅｓｃｒｉｐｔｏｒ ) for P≧1 voxels (v ｃ ) and using it to predict the location of the center (11) of the cochlea (10) in said image (M); A method comprising:

2. The cochlear descriptor image (E ｄｅｓｃｒｉｐｔｏｒ ) forming the first, second and third gradient accumulation images (E １ , E ２ , E ３ ) Hadamard product (E １ ○ E ２ ○ E ３ ) ｉｊ = (E １ ) ｉｊ (E ２ ) ｉｊ (E ３ ) ｉｊ The method of claim 1 , corresponding to calculating:

3. The respective gradient lines are aligned at the respective starting points (p start,１ , p start,２ , p start,３ ) to each of the end points (p ｅｎｄ,１ , p ｅｎｄ,２ , p ｅｎｄ,３ 3. The method of claim 1 or claim 2, wherein the length of each first gradient line is greater than the length of each third gradient line, and the length of each third gradient line is greater than the length of each second gradient line.

4. The first, second and third starting points p start,１ , p start,２ , p start,３ is calculated according to p start,ｉ =position(v)+start ｉ ・g i=1、2、3 The first, second and third end points p ｅｎｄ,１ , p ｅｎｄ,２ , p ｅｎｄ,３ , is calculated according to p ｅｎｄ,ｉ =position(v)+end ｉ ・g i=1、2、3 where position(v) denotes the position of the respective voxel (v) in the image (M), - start １ is in the range of 2.7mm to 3.3mm, especially the start １ is equal to 3.0 mm, - start ２ ranges from -3.3mm to -2.7mm, especially start １ is equal to -3.0 mm, - start ３ is in the range of 0.45mm to 0.55mm, especially start ３ is equal to 0.5 mm, - end １ is in the range of 4.5mm to 5.5mm, especially end １ is equal to 5.0 mm, - end ２ is in the range of -1.8mm to -2.2mm, especially end ２ is equal to -2.0 mm, - end ３ is in the range of 1.8mm to 2.2mm, especially end ３ is equal to 2.0 mm, The method according to any one of claims 1 to 3.

5. providing a statistical shape model (SSM) of an ear including at least a cochlear portion, said cochlear portion corresponding to said cochlea (10) of said ear; The method of any one of claims 1 to 4, further comprising:

6. For each voxel of said number P of voxels (v ｃ ) Each of the voxels (v ｃ defining a volume (B1) in said image (M) centered on segmenting the cochlea (10) within said volume (B1) using a Hessian-based enhancement filter, thereby obtaining a segmented cochlear region of said image (M), and generating a 3D surface of the cochlea from said segmented cochlear region, in particular by using a marching cubes algorithm; Fitting the cochlear portion of the statistical shape model (SSM) to the 3D surface; calculating a surface similarity metric indicating the similarity between the surface of the cochlear portion of the statistical shape model (SSM) and the 3D surface, - the number P of voxels (v ｃ ), the surface similarity index satisfies a predefined criterion. ｃ ) as the center (11) of the cochlea (10), and in particular the first voxel (v ｃ ) is selected as the center of the cochlea, The method of claim 5 further comprising:

7. - fitting the cochlea portion of the statistical shape model (SMM) to the segmented cochlea region using the determined center (11) of the cochlea (10), the statistical shape model (SSM) further comprising a semicircular canal portion corresponding to the three semicircular canals (12) of the ear; - predicting the positions of the three semicircular canals (12) in the image (M) as the positions of the three semicircular canals (12) in the semicircular canal portion of the statistical shape model (SSM); - defining a volume (B2) containing said predicted position, - segmenting said three semicircular canals (12) in said volume (B2), thereby obtaining segmented semicircular canal regions of said image (M); The method of claim 6 further comprising:

8. - fitting the cochlea and semicircular canal portions of the statistical shape model (SSM) to the segmented cochlea and semicircular canal regions; The method of claim 7, further comprising:

9. - predicting the positions of the incus (5), malleus (6) and stapes (7) in said image (M) as said positions of the incus (5), malleus (6) and stapes (7) in the ossicular region of said statistical shape model (SSM), said ossicular region of said statistical shape model corresponding to the incus (5), malleus (6) and stapes (7) of said ear, - defining a volume (B4) containing said predicted position, - segmenting said incus, malleus and stapes in said volume, thereby obtaining segmented ossicular regions of said image (M) corresponding to the incus, malleus and stapes in said image (M); The method of claim 8 , further comprising:

10. - predicting the location of said tympanic membrane (13) as said location of said tympanic membrane (13) in a tympanic membrane part of said statistical shape model (SSM), said tympanic membrane part of said statistical shape model corresponding to said tympanic membrane (13) of said ear, - defining a volume (B3) containing said predicted position, - segmenting said tympanic membrane (13) within said volume (B3), thereby obtaining a segmented tympanic membrane region of said image (M); The method of claim 8 or 9, further comprising:

11. - predicting the area, orientation and boundaries of the location of the ear canal (14) based on the segmented eardrum region of the image (M); - defining a volume (B5) containing said predicted area, - segmenting said ear canal (14) within said volume (B5), thereby obtaining a segmented ear canal region of said image (M); moreover, - predicting the position and orientation of said outer surface (15) of said temporal bone (16) based on said segmented ear canal region of said image (M); - defining a volume (B6) containing said predicted position, - segmenting said temporal bone (16) in said volume (B6) using a thresholding process, thereby obtaining a segmented temporal bone region of said image (M) and detecting said outer surface using said orientation of said ear canal (14); The method of claim 10 further comprising:

12. - predicting the location of the inner ear canal (17) based on the segmented cochlear region in the image; - defining a volume (B7) containing said predicted position, - segmenting said ear canal (17) within said volume (B7), thereby obtaining a segmented ear canal region of said image (M); The method according to claim 6 or any one of claims 7 to 11 based on claim 7, further comprising:

13. - predicting the location of the facial nerve (18) based on the segmented cochlea region, the segmented semicircular canal region, and the segmented ear canal region of the image (M); - defining a volume (B8) containing said predicted position, - segmenting said facial nerve (18) within said volume (B8), thereby obtaining a segmented facial nerve region of said image (M); The method of claims 6, 8 and 11, further comprising:

14. 14. The method of claim 6, further comprising: graphically visualizing for a user on a display (102) at least one of the segmented cochlea region, the segmented semicircular canal region, the segmented ossicular region, the segmented tympanic membrane region, the segmented external ear canal region, the segmented temporal bone region, the segmented internal ear canal region, and the segmented facial nerve region.

15. A system (100) for automatically segmenting a human cochlea (10) contained in an input image (M) composed of voxels (v), each voxel including an intensity value indicative of the intensity of the respective voxel, said system (100) comprising at least one processor (101) and a display (102), wherein the at least one processor (101) and the display (102) are adapted to perform the method according to any one of claims 1 to 14.

Images