Eur Arch Otorhinolaryngol (2009) 266:507–518 DOI 10.1007/s00405-008-0777-7

RHINOLOGY

CT-based manual segmentation and evaluation of paranasal sinuses S. Pirner · K. TingelhoV · I. Wagner · R. Westphal · M. Rilk · F. M. Wahl · F. Bootz · Klaus W. G. Eichhorn

Received: 31 May 2008 / Accepted: 16 July 2008 / Published online: 21 August 2008 © Springer-Verlag 2008

Abstract Manual segmentation of computed tomography (CT) datasets was performed for robot-assisted endoscope movement during functional endoscopic sinus surgery (FESS). Segmented 3D models are needed for the robots’ workspace deWnition. A total of 50 preselected CT datasets were each segmented in 150–200 coronal slices with 24 landmarks being set. Three diVerent colors for segmentation represent diverse risk areas. Extension and volumetric measurements were performed. Three-dimensional reconstruction was generated after segmentation. Manual segmentation took 8–10 h for each CT dataset. The mean volumes were: right maxillary sinus 17.4 cm³, left side 17.9 cm³, right frontal sinus 4.2 cm³, left side 4.0 cm³, total frontal sinuses 7.9 cm³, sphenoid sinus right side 5.3 cm³, left side 5.5 cm³, total sphenoid sinus volume 11.2 cm³. Our manually segmented 3D-models present the patient’s individual anatomy with a special focus on structures in danger according to the diverse colored risk areas. For safe robot assistance, the high-accuracy models represent an average of the population for anatomical variations, extension and volumetric measurements. They can be used as a database

S. Pirner · K. TingelhoV · I. Wagner · F. Bootz · K. W. G. Eichhorn Clinic und Policlinic of Otolaryngology/Ear, Nose and Throat Surgery, University of Bonn, Bonn, Germany e-mail: [email protected] K. W. G. Eichhorn (&) Universitätsklinikum Bonn, Sigmund-Freud-Str. 25, 53127 Bonn, Germany e-mail: [email protected] R. Westphal · M. Rilk · F. M. Wahl Institute of Robotics and Process Control, Technical University of Braunschweig, Braunschweig, Germany

for automatic model-based segmentation. None of the segmentation methods so far described provide risk segmentation. The robot’s maximum distance to the segmented border can be adjusted according to the diVerently colored areas. Keywords Manual segmentation · Paranasal sinuses · FESS · Computed tomography · 3D model · Robot

Introduction Functional endoscopic sinus surgery (FESS) has become the standard treatment for chronic paranasal sinus pathology. In recent years computer-assisted surgery (CAS) has been developed for improving FESS. The project “Robotassisted intuitive endoscope navigation in endonasal surgery using preoperative computed tomography (CT) or magnetic resonance imaging (MRI) analysis” uses automatic robot assistance technique in guiding the endoscope during FESS. Surgical workXow analyses demonstrate the advantage of this aim [1]. The robot’s workspace has to be deWned preoperatively with the help of manually segmented CT data. Manual segmentation results can be used for trajectory planning, for surgical training as well as for navigation during FESS. The FESS technique aims to restore normal physiology by reestablishing normal drainage and ventilation of the sinuses [2]. Due to restricted access to the paranasal sinuses, the technique has its limitations in reduced workspace. The FESS surgeon needs substantial experience and training. The disadvantage of conventional FESS for the surgeon is the restriction to one hand for manipulation while the other hand is holding the endoscope. Robotic assistance in endoscopic guidance will reduce time and

123

508

increase security during FESS [1]. Surgical manipulation close to the base of the skull, the internal carotid artery or the optic nerve is critical. Common complications are cerebrospinal Xuid leak, decreased sense of smell, blindness or even permanent disability and death because of carotid artery injury and other incidence of bleeding [3]. The FESS surgeon has to acquire specialized skills and knowledge of the anatomy of nasal cavity and paranasal sinuses and variations to avoid complications. Anatomical abnormalities that may predispose the patient to chronic sinusitis need to be detected. The relationships between anatomical variations, like septal deviations, Haller cells, Agger nasi cells or Onodi cells, and recurrent diseases are controversial [2–5] but these conditions must be considered before surgery and integrated in the preoperative planning. The extent of the sinuses procedure has to be deWned preoperatively. It is important to be aware of sinonasal diseases mimicking inXammatory patterns of sinusitis, especially in long-lasting cases with lack of response to medical treatment. In severe unilateral disease, a neoplastic process should always be ruled out. A CT dataset and its preoperative analysis is mandatory to identify the patient’s individual anatomy and to avoid complications during FESS [6]. In complex anatomy, orientation during FESS can be diYcult. To play it safe, CAS and navigation based on CT data during FESS are recommended [7, 8]. Computed tomography-based segmentation is useful for evaluation of nasal cavities and paranasal sinuses. Volume measurement is the simplest and most important index for analyzing paranasal sinuses [9]. Manual segmentation needs high user expertise and is a time-consuming process due to the per-slice user-interaction required [10]. Nevertheless, it ensures very high accuracy and quality for developing automatic segmentation. Manual segmentation of the endonasal cavities and 3D models can be used for computer and for robotic assistance during FESS.

Materials and methods A total of 50 computed tomography datasets were generated by the Institute for Radiology, University of Bonn. The database of the paranasal sinuses was obtained between 5 December 2003 and 2 May 2006. CT was performed using a Philips 16 multislice spiral CT. CT datasets had a slice thickness between 1 and 2 mm. The pixel spacing was between 0.6 mm £ 0.6 mm and 0.3 mm £ 0.3 mm. CT datasets were obtained by patients with a clinical symptom presumably referrable to the paranasal sinus region. The CT datasets were preselected to achieve a standard size for the development of automatic segmentation. Patient’s histories were not taken into account. The paranasal sinuses’

123

Eur Arch Otorhinolaryngol (2009) 266:507–518

selected had to be fully developed, completely visible and of normal anatomical formation. We only included patients older than 16 without dysplastic midface and chose a balanced male to female ratio. CT studies were eliminated which involved either prior surgery or a tumor which destroyed some or most of the anatomical structures being studied and segmented. CT datasets with a slice thickness of no more than 2 mm ensured high-quality results. We obtained CT datasets from 24 male and 23 female patients and 3 cadaver heads. The average age of the patients was 39 years (range 16–78 years). All the 50 CT data were used for extension and volumetric measurement. A total of 47 CT datasets were additionally evaluated for pathological diagnoses and anatomical variations. The three cadaver heads showed alterations of the soft tissue due to postmortal eVects and therefore were excluded for pathological evaluation. We examined the CT datasets for any pathology in paranasal sinuses. We considered the prevalence of Agger nasi cells, Haller cells, Onodi cells and concha bullosa. DeWnitions were used according to Bolger et al. [11] and Kantarci et al. [12]. A concha bullosa was considered to be any pneumatization of the middle turbinate visible in coronal CT dataset. Agger nasi cells accounted the most anterior ethmoid cells that are located anterior, lateral and inferior to the frontal recess. Haller cells were regarded as the ethmoidal cells that develop into the Xoor of orbit adjacent to and above the maxillary sinus ostium. Onodi cells were considered to be posterior ethmoid cells which pneumatized far laterally and to some degree above the sphenoid sinus. Our self-developed software enables loading and displaying DICOM datasets. The CT datasets can be examined in coronal, axial and sagittal views layer by layer. The software provides line segmentation. The user marks several points and the software draws straight lines between two points. Three diVerent colors can be chosen for segmentation. The user can set landmarks. It is possible to measure the distance between two landmarks to size the nasal cavity and the paranasal sinuses in all directions. It is possible to change HoundsWeld windows from bone to mucosa visualization or manually, as required. Volumetric measurement is possible after segmentation, which can be achieved with region growing. A total of 50 CT datasets were segmented manually layer by layer in 150–200 coronal planes. For segmentation we used line segmentation. All paranasal sinuses including the nasal cavity were outlined according to its inner mucosa surface. The mucosa was considered to be inside the segmented region, only bony structures were marked. We did not mark either the ethmoid cells individually or the turbinates. Figure 1 shows a manually segmented CT dataset.

Eur Arch Otorhinolaryngol (2009) 266:507–518

509

Fig. 1 Screenshot of the manual segmentation software

Three diVerent colors were chosen for segmentation. They represent the borders to areas with diVerent risk, due to the complications described during FESS or contiguousness to potential risk areas. The areas close to orbit and skull base were classiWed as high-risk regions and marked red. Lower risk regions were marked blue. Structures removable during FESS, like the uncinate process or the maxillary sinus ostia, were marked green. Figure 2 shows a diVerently colored manual segmentation. A total of 1,200 landmarks were set, 24 for each CT dataset. The landmarks were bony structures which were recognizable in all CT datasets, reproducible, small-sized and located in diVerent planes. Landmarks will be used for CT registration. We chose for example both sides of the orbit’s centroid, infraorbital foramina, anterior and posterior nasal spine, crista galli and styloid process. Three-dimensional reconstruction was generated after segmentation results were postprocessed. We used a Gaussian Wlter Wve times consecutively in order to smooth the binary segmentation results. Marching cubes were then used for generating the 3D mesh. For visualization, we used OpenGL which provides the visualization of the model in 62.5 frames per second. It ensures a frictionless movement in all directions. The mesh consists of vertexes which are represented by the corresponding coordinates on the x-, y-, and z-axes. Each three vertexes form a triangle. The 3D model can be stored as obj-Wle. All vertexes with the corresponding coordinates and all triangles with the corresponding

Fig. 2 Example of a risk-orientated manual segmentation. The areas close to orbit or skull base were marked red, lower risk regions were marked blue and parts removable during Functional Endoscopic Sinus Surgery (FESS) were marked green. The mucosa was inside the segmented region, only bony structures were marked

vertexes are listed and stored in the text Wle. The 3D models represent the colored regions at risk. They are visible without the CT dataset’s background and rotatable in all directions. It is possible to select an explosion 3D-model, which makes is possible to look inside the segmented 3Dmodel.

123

510

The database selected has to be representative of the average in the population. That is why we evaluated our CT datasets for pathological diagnoses, anatomical variations, extension and volumetric measurements. The values were evaluated separately for men and women. To evaluate the extension of the paranasal sinuses we measured height £ width £ length in mm (length: anterior–posterior; height: cranio-caudal; width: transverse). We studied all paranasal sinuses, including nasal cavity en bloc and each sinus separately. Frontal sinus and sphenoid sinus were measured each as one cavity, because the dividing septum is not always existent or may be inclined. We did not subdivide into right and left sides. The distance between two landmarks was computed in x-, y-, and z-direction parallel to the coordinate system of the CT dataset. The maximum distance in the chosen direction was thus measured. We did not restrict ourselves to one CT slice for all measurements. The distance was marked from bone to bone. The volumes of maxillary, frontal and sphenoid sinuses were determined semi-automatically. First a volume of interest was deWned. Based on the gray level of a seed point chosen manually, all voxels with a similarly deWned gray value were added to the segmented region (region growing). This method was not suitable for pathological sinuses like modiWed mucosa or sinusitis. In this case all the voxels inside an enclosed manually segmented area were calculated as shown in Fig. 3. Therefore, the volumetric measurement had to be stopped with the Wrst and the last coronal plane slice which had been selected. For volume computation, the number of voxels belonging to each sinus was counted and the result was multiplied by the volume of each voxel in cm³. For the sphenoid sinus and frontal sinus we measured the volume separately for right and left cavities. The values of the right and left sides were added to

Fig. 3 Volumetric measurement of paranasal sinuses

123

Eur Arch Otorhinolaryngol (2009) 266:507–518

obtain total volumes. The values are sub-divided between right sides, left sides and total volumes, which implies right and left side together including unsepted sinuses. This was done to expound an irregular division of the sphenoid or frontal septum. DiVerences between genders concerning extension and volumetric measurements were analyzed using the unpaired t test. All reported signiWcance levels are considered as P < 0.05, high signiWcance is considered as P < 0.001.

Results Manual segmentation took on average 8–10 h per CT dataset. A total of 50 3D-models of the nasal cavity and the paranasal sinuses were computed. The 3D-mesh, e.g., in Fig. 4a, consists of 78,685 vertexes. Pictures in an upright position of the models received are displayed in Fig. 4 from a front perspective. They show the diverse colored risk areas. Risky regions are marked red. They are located close to areas like the base of skull, the brain or the orbit. This explains why mainly the upper parts are marked red (Fig. 5c). Borders close to areas of lower risk are marked blue. These are for example mostly lower parts of the 3D models, the nasal septum and the ventral part of the frontal sinuses. Structures marked green can be removed during FESS, e.g., the uncinate process. They are located inside the nasal cavity and therefore not visible in Figs. 4 and 5. The 3D models illustrate the individual anatomical appearance of paranasal sinuses. The shape of maxillary sinuses pictured has a similar appearance in Fig. 4. The pictured frontal sinuses shown are diversely shaped. Figure 4a demonstrates aplastic, 4b asymmetric and 4c distinct shaping of frontal sinuses. Three-dimensional models are rotatable in all directions. This is demonstrated by illustrations (Fig. 5) taken from a moving model. Figure 5a demonstrates the model from the left, 5b from behind and 5c from above. The diVerent risk areas can be examined separately (Fig. 6c). It is also possible to compute an explosion 3D model, as demonstrated by Fig. 6. Therefore, the border marked red is pulled upwards. This visualization makes it possible to look inside the computed 3D model of paranasal sinuses and nasal cavity. The removable parts during FESS in explosion models, which are marked green, are especially noticeable. Figure 6a demonstrates a front view, 6b a view from the side. Figure 6c represents visualization without blue marked borders from behind. Segmented 3D models based on CT datasets demonstrate the patient’s individual anatomy. Nevertheless, only a few anatomical variations which can cause recurrent sinusitis are visible. Nasal septal deviations and the constitution of the uncinate process are illustrated. We evaluated the CT

Eur Arch Otorhinolaryngol (2009) 266:507–518

511

Fig. 4 3D models with diverse colored risk areas demonstrate the individual anatomical appearance of paranasal sinuses. Frontal sinuses can be aplastic, asymmetric or in a distinct shape

Fig. 5 The 3D models are rotatable. a The model pictured from the left, b from behind and c from above

Fig. 6 Visualization of an explosion 3D model from diVerent perspectives. a A front view, b a view from the side. c Visualization without blue segmentation from behind

datasets for additional anatomical variations and pathological diagnoses. We examined in 40.4% ethmoid sinusitis, in 31.9% maxillary sinusitis, in 27.7% maxillary sinus cysts, in 8.5% sphenoid sinusitis and in 4.3% frontal sinusitis. Septal deviation occurs in 57.5%, concha bullosa in 34.0%. Agger nasi cells are present in 57.4%, Haller cells in 8.5% and Onodi cells in 2.1%. The anatomical variations of our patients are within the range of measurements obtained by other studies (Fig. 7). Mean extensions for all the paranasal sinuses including the nasal cavity en bloc are 84.0 mm £ 81.5 mm £

104.6 mm, for the right maxillary sinus 39.4 mm £ 29.7 mm £ 40.7 mm, for the left maxillary sinus 39.6 mm £ 28.9 mm £ 40.6 mm, for the frontal sinus 28.7 mm £ 54.3 mm £ 21.4 mm and for the sphenoid sinus 25.8 mm £ 36.6 mm £ 32.5 mm. All measurements of paranasal sinuses including the nasal cavity en bloc are signiWcantly larger in male than in female patients. Maximum height is larger for men than for women (P = 0.011). The width measured diVers signiWcantly between men and women (P = 0.036). Maximum length is even very signiWcantly larger for men than for women (P = 0.001). The

123

512

Eur Arch Otorhinolaryngol (2009) 266:507–518

Fig. 7 Anatomical variations in comparison with other authors

length of the frontal sinuses also diVers very signiWcantly between men and women (P = 0.001). The height of maxillary sinuses on the right sides is signiWcantly higher in male patients (P = 0.037). All the results of extension measurements including the sub-division into men and women are listed in Table 1. The expansion measurements are larger or in a higher range than other research groups reported, as shown in Table 2. If no division between right and left side is given by the author, we decided to compare the results with the right side.

In our measurements the maxillary sinus volume ranges from 4.4 to 31.8 cm³. The mean volume of the right maxillary sinus is 17.4 cm³, on the left side 17.9 cm³. Three patients do not have a septum between right and left sphenoid sinuses. In these cases we measured the complete volumes. Four patients have nearly aplastic or asymmetric frontal sinuses. In one of these small-sized frontal sinuses we could only measure the complete volume. The mean volume of the right frontal sinus is 4.2 cm³ and on the left side 4.0 cm³. Total frontal sinuses are on average 7.9 cm³, including unsepted sinuses. The mean volume of the sphenoid

Table 1 Extension measurements of the paranasal sinuses and nasal cavity Paranasal sinuses

Estimated 3D measurements (mm) Height § SD (min–max)

Width § SD (min–max)

Length § SD (min–max)

39.4 § 5.0 (23.0–50.0)

29.7 § 4.0 (19.5–37.9)

40.7 § 3.6 (29.0–47.7)

Maxillary sinus R L

39.6 § 4.6 (23.0–51.0)

28.9 § 3.9 (17.4–6.3)

40.6 § 3.7 (25.3–47.7)

Frontal sinus

28.7 § 10.3 (5.0–48.0)

54.3 § 16.7 (12.3–92.5)

21.4 § 7.8 (5.4–39.2)

Sphenoid sinus

25.8 § 4.1 (16.0–38.0)

36.6 § 6.4 (26.6–49.1)

32.5 § 5.0 (17.6–42.5)

Maximum extension

84.0 § 13.1 (52.0–108.0)

81.5 § 11.0 (23.6–97.4)

104.6 § 8.5 (87.3–125.4)

R

38.0 § 4.5

29.2 § 3.7

40.0 § 3.0

L

38.6 § 3.5

28.9 § 3.0

40.1 § 2.8

Frontal sinus

26.3 § 9.7

49.5 § 16.6

17.6 § 5.2

Sphenoid sinus

24.8 § 4.4

35.5 § 6.9

32.0 § 4.8

Maximum extension

78.3 § 11.4

78.2 § 13.3

101.4 § 7.6

41.3 § 5.1

30.6 § 4.1

42.0 § 3.6

Women Maxillary sinus

Men Maxillary sinus R L

40.5 § 5.4

29.5 § 4.5

41.2 § 4.4

Frontal sinus

30.7 § 10.8

58.0 § 15.7

24.9 § 8.2

Sphenoid sinus

27.3 § 3.3

37.5 § 5.8

34.1 § 3.8

Maximum extension

89.3 § 12.8

85.2 § 6.6

109.6 § 6.3

R right, L left; Length anterior–posterior, height cranio-caudal, width transverse

123

Eur Arch Otorhinolaryngol (2009) 266:507–518 Table 2 Extension measurement of paranasal sinuses in comparison with other authors

Paranasal sinus

513

References

Average (mm) Right

Left

Uchida et al. [25]

35.7

33.4

Barghouth et al. [23] (>16 years)

36.3 § 6.2

Total

Maxillary sinus Height

Width

Length

Own measurement

39.4 § 5.0

Uchida et al. [25]

25.9

Ariji et al. [22]

27.0 § 6

Barghouth et al. [23] (>16 years)

27.5 § 4.2

Own measurement

29.7 § 4.0

Spaeth et al. [24]

32.3 M 29.3 F

Uchida et al. [25]

30.3

Ariji et al. [22]

35.6 § 74.7

Barghouth et al. [23] (>16 years)

38.8 § 3.5

Own measurement

40.7 § 3.6

Spaeth et al. [24]

41.0 M 38.2 F

39.6 § 4.6

29.9

40.6 § 3.7

Frontal sinus Height

Own measurement Barghouth et al. [23] (>16 years)

Width

Length

28.7 § 10.3 21.9 § 8.4

Own measurement

54.3 § 16.7

Barghouth et al. [23] (>16 years)

24.5 § 13.3

Spaeth et al. [24]

28.0 § 7.1 M 26.4 § 6.7 F

Own measurement

21.4 § 7.8

Barghouth et al. [23] (>16 years)

12.8 § 5.0

Spaeth et al. [24]

17.4 § 5.2 M 16.1 § 5.8 F

Sphenoid sinus Height

Own measurement Barghouth et al. [23] (>16 years)

Width

R right, L left, Total whole volumes measured without division into right and left sinuses, height cranio-caudal, width transverse, length anterior–posterior, M male, F female

25.8 § 4.1 22.6 § 5.8

Barghouth et al. [23] (>16 years)

12.8 § 3.1

Spaeth et al. [24]

33.5 § 6.1 M 32.8 § 7.1 F

Own measurement Length

Barghouth et al. [23] (>16 years)

36.6 § 6.4 23 § 4.5

Own measurement Spaeth et al. [24]

sinus is on the right side 5.3 cm³, 5.5 cm³ on the left side and 11.2 cm³ for the total volume, including unsepted sinuses. The mean volumes for maxillary, frontal and sphenoid sinuses sub-divided between males and females are given in Fig. 8. All results of volumetric measurements are included in Table 3. The comparison between female and male volumes revealed no signiWcant diVerences for maxillary sinuses. The volumes of frontal sinuses are signiWcantly larger in male than in female patients. There is a diVerence between men and women for the right frontal sinuses (P = 0.007),

21.4 § 7.8 32.9 § 6.5 M 28.0 § 5.7 F

the left frontal sinuses (P = 0.010) and for the whole frontal sinuses (P = 0.004). The volumes of the sphenoid sinuses are signiWcantly larger in men than in women for the left sides (P = 0.016) and for the whole volumes of sphenoid sinuses (P = 0.002). The volumetric values measured do not diVer from those obtained by other studies, as shown in Table 3. The reported results are sub-divided between right sides, left sides and total volumes, which implies right and left side together including unsepted sinuses. Where available there is a sub-division between female and male values. Values

123

514

Eur Arch Otorhinolaryngol (2009) 266:507–518

Fig. 8 Volumes of the Paranasal sinuses subdivided between genders in cm³. R right side, L left side, Max maxillary, Fron frontal, Sphen sphenoid

in italics represent volumetric measurements calculated by directly injecting a material. Non-italic values imply the results of volumes measured by CT-based segmentation.

Discussion Segmentation of paranasal sinuses is of great interest for surgical ear, nose and throat (ENT) workXow. CAS and navigation systems based on CT datasets are recommended in complex anatomical situations [7, 8]. For using CAS, additional image-based information is needed. We have evaluated an approach to obtain the individual anatomy of the nasal cavity and the paranasal sinuses in manually segmented 3D models. The usefulness of three-dimensional CT models is demonstrated in other complex anatomical regions [13]. None of the manual and semi-automatic segmentation methods so far described provides risk segmentation. We consider the risk-orientated segmentation to be highly useful. Not being penetrated, the robot’s maximum distance to the segmented border can be adjusted according to the diVerently colored areas. Additionally to these diVerent areas, other measurements can be included. The results of force data and tissue elasticity obtained during FESS [7, 14] may even improve our chosen risk areas. It is well known that the mucosal thickness can change over time in between the CT data processing date and performed FESS, as it undergoes cyclical changes. Our method obtains a segmentation of bones and is therefore less fragile for mucosal variations. Our segmentations can be used for virtual endoscopy and surgical training, too. The novice surgeon can learn surgical treatment with a special focus on structures at risk with our risk-orientated models. Complex anatomy, especially, can be visualized clearly. There are three possibilities for segmentation: manual, semi-automatic and automatic segmentation. Automatic segmentation aims at reducing interaction time and enhancing precision.

123

There are some research groups using manual segmentation. Strauss et al. [7] segmented a workspace manually for evaluating virtual assistance in endoscope guiding. The time taken for segmentation was 15 min. Borders were drawn manually in axial CT slices, projected into the next slice and adapted there. Our method provides higher accuracy but needs more user interaction, due to strictly manual interaction. A Computed tomography based workspace for using powered instruments has been deWned [15]. High security is necessary for employing invasive instruments also known as microdebriders or soft tissue shavers. These are automatically stopped by mechanotronical assistance if critical structures are reached (navigated control). A segmentation method based on CT data by manually setting vertices which were chained was described. Borders close to critical areas have to be segmented accurately to ensure the shavers automatically shut down outside the segmented regions. Nevertheless, the reported segmentation is imprecise. Only a few points were set and chained automatically. The time needed for segmentation was described as being less than 5 min. Our models are far more accurate due to performed per-slice user interaction and deWned risk areas. They may be better appropriated for using invasive instruments. Our manual segmentation of nasal cavity and paranasal sinuses took on average 8–10 h per CT dataset. Doing this preoperatively for every patient undergoing FESS is too time consuming in everyday workXow. Automatic segmentation is needed to reduce interaction time. Currently there are only semi-automatic methods for paranasal sinuses segmentation available. They are mostly based on 3D intensity-based region growing or watershedtransformation. Region Growing merges neighboring voxels that meet a speciWed intensity threshold criterion. Watershed-transformation divides regions which diVer distinctly, for example in their gray level. For the segmentation of simple structured objects, like the carotid artery [16] or multiple sclerosis lesions [17], semi-automatic methods are expedient. The anatomy of the paranasal sinuses is far more complex. The gray levels cannot be deWned as easily,

Eur Arch Otorhinolaryngol (2009) 266:507–518 Table 3 Measured volumes of the paranasal sinuses in comparison with other authors

Paranasal sinuses

515

References

Estimated volume Average (cm3)

Maxillary sinus

Frontal sinus

Right

Left 8.6 § 5.4

Schumacher et al. [26]

9.4 § 5.1

Uchida et al. [25]

11.3 § 4.6

Anagnostopoulous et al. [27]

11.6

11.9

Karakas et al. [28] (>25 years)

15.0 § 5.2 M 11.1 § 4.5 F

16.0 § 6.7 M 11.5 § 5.5 F

Own measurement

17.4 § 5.9 19.5 § 6.2 M 15.9 § 5.1 F

17.9 § 5.5 19.2 § 6.6 M 16.8 § 4.2 F

Emirzeoglu et al. [29]

18.0 § 6.0 19.8 § 6.3 M 16.0 § 5.0 F

Barghouth et al. [23] (>16 years)

18.3

Kawarai et al. [9]

21.2 § 6.5

23.0 § 6.7

Schumacher et al. [26]

3.1 § 2.3

4.6 § 3.6

Karakas et al. [28] (> 25 years) Own measurement

8.4 § 4.0 M 3.5 § 2.4 F 4.2 § 3.5 5.4 § 3.9 M 2.8 § 2.2 F

4.0 § 2.6 4.9 § 3.0 M 3.0 § 1.9 F

Kawarai et al. [9]

Emirzeoglu et al. [29]

Sphenoid sinus

Italics volumes calculated by directly injecting a material, Nonitalics volumes calculated by manually segmented CT data

5.8 § 4.1 7.5 § 4.3 M 4.1 § 2.9 F

Barghouth et al. [23] (>16 years)

2.7

Schumacher et al. [26]

3.2 § 2.5

11.6 § 0.8

3.5 § 2.8 8.5 § 4.2 M 7.9 § 3.0 F

Own measurement

5.3 § 3.5 6.2 § 3.7 M 4.6 § 3.0 F

Emirzeoglu et al. [29]

6.9 § 3.7 7.7 § 4.0 M 6.1 § 3.2 F

Kawarai et al. [9]

due to their non-homogeneous constitution of air, bone and mucosa. In these cases semi-automatic segmentations are limited. The problems of using semi-automatic segmentations in more complicated areas are reported on by several authors. Salah et al. [10] evaluated a semi-automatic CTbased segmentation of paranasal sinuses using 3D region growing. Apelt et al. [18] used the watershed-transformation. Both methods distinctly reduced time needed for segmentation, but did not achieve full accuracy [10, 18]. The results had to be reworked manually, especially in cases where critical structures were not respected. Structures

7.9 § 5.5 10.2 § 6.2 M 5.5 § 3.5 F 8.1 § 5.1 11.6 § 4.2 M 4.6 § 3.2 F

Karakas et al. [28] (>25 years)

M male, F female, Total total volume including unsepted sinuses

Total

5.5 § 3.0 6.5 § 3.1 M 4.6 § 2.6 F

11.2 § 4.5 13.3 § 4.4 M 9.6 § 4.0 F 13.6 § 0.7

15.4 § 6.9 17.1 § 7.4 M 13.7 § 6.2 F

which do not diVer distinctly in their contrast need more user interaction. Semi-automatic segmentation is still very time consuming and not as accurate as manual segmentation and therefore not feasible in daily workXow. Automatic segmentation is needed [19] and developed in our research group. Depending on the complex anatomy, a model-based approach seems to be necessary, even though there is a high interindividual variability [20]. The database used as a model has to be precise for this method, because the automatic version depends on it. The model-based segmentation is meant to be able to segment many patients’ CT

123

516

datasets automatically. Therefore, the database selected has to be representative of the average in the population. We used the manual segmentations presented as an origin. Landmarks were set in each CT dataset and are used for CT registration. Nevertheless, the model-based method has its limitations, because it relies on a database. For example, our automatic model-based segmentations will have problems with segmenting structures which are destroyed because of tumor growth. We eliminated CT datasets for manual segmentation when they involved a tumor, which destroyed some or most of the anatomic structures being studied and segmented. For a safe robot-assisted movement, we wanted to ensure a representative average of the population by evaluating our CT datasets for pathological diagnoses, anatomical variations, extension and volumetric measurements. It is also important to know soft tissue and bone properties of nasal cavity and paranasal sinuses for a safe robot-assisted endoscope movement during FESS [14]. First of all, we evaluated pathological diagnoses. We examined any paranasal sinusitis or maxillary cyst, septal deviation, concha bullosa or ethmoidal cells as Agger nasi cells, Haller cells and Onodi cells. We opted for the abovementioned anatomical variations, due to their potential clinical relevance. Our data are all within the range of values obtained by other studies, which are shown in Fig. 7, even though all CT datasets were preselected [4, 5, 11, 12, 21]. Second, we evaluated anatomical variations. The relationship between anatomical variations and recurrent disease are controversial [2–4]. The wide range of pathological diagnoses described has been reported by several authors [4, 5, 11, 12]. This wide distribution may be explained by the variations in descriptions and deWnitions for pneumatization and localization of ethmoid cells given by several authors. We opted for the deWnitions used by Bolger et al. [11] and Kantarci et al. [12]. Their deWnitions seemed to be coherent. Anatomical variations should be considered as potential predisposing factors in sinusitis and not as the underlying etiology of the disease, as other authors reported [3]. Next, we measured the extension of the paranasal sinuses for further evaluation. Some research groups measured their values on one CT slice perpendicular to deWned lines [22]. They may have chosen this restriction because they wanted to achieve reproducible values. To achieve maximum values we did not restrict ourselves to one CT slice. Barghouth et al. [23] even measured some values slightly oblique for this purpose. In our case, reproducibility is ensured by measuring in x-, y-, and z-direction parallel to the coordinate system of the CT dataset. We consider it to be important to consider expansions of all three dimensions, as Barghouth et al. [23] also reported. Most reported extension values were less than ours [22, 25].

123

Eur Arch Otorhinolaryngol (2009) 266:507–518

Variations in deWnitions, inherent diVerences in study population and especially diVerences in analyzing methods can bring about other results. Uchida et al. [25] for example used impression material for volumetric measurements of maxillary sinuses. They took the extension measurements from the casts they produced. Another reason for diVering extension values can be the method used for measuring frontal and sphenoid sinuses. It was not always reported if a sub-division between the right and left side was performed. Barghouth et al. [23] measured the right and left side of sphenoid sinuses even though he could not visualize the septum in 60%. In these cases they reported the half of the total volume measured. The values listed in Table 2 show that most authors used this method. This explains why especially the values given for the width are diVerent from our measurements. We did not sub-divide into right and left frontal and sphenoid sinuses for extension measurements, due to the location of the septum being often irregular or not visible. Our extension measurements also reveal great variability in frontal sinuses expansions which is demonstrated by large standard deviations and high minimum to maximum values. This is conWrmed by other results [23]. The comparison between male and female values shows signiWcant diVerences for the extension of all the paranasal sinuses en bloc. This is important for the robotic trajectory planning. Furthermore, volumetric measurements were performed, as this is the simplest and most important index for analyzing paranasal sinuses [9]. We used our manually segmented and computed 3D models, which ensured high accuracy. Especially methods which approximate the volume by the use of rotating forms like ellipses are imprecise. This may be one reason why Barghouth et al. [23] obtained especially small values for sphenoid sinuses. The second reason could be that they used children aged from 16 to 17. In comparison with other authors using adult age groups, this may explain their smaller values as well. In general, the methods of rotating forms only seem expedient for less complex structures. Our results for volumetric measurements were in range with other authors. There were obvious diVerences in the values obtained, depending on the method used. Some authors measured the volume by directly injecting a material [25–27]. This procedure cannot be used in living subjects. Computed tomography was employed by several authors for volume measuring [9, 28, 29]. Emirzeoglu et al. [29] also used CT datasets and therefore are added to this group. They performed measurements by a square test grid system, which was superimposed on the sinuses. Sections making contact with the sinus of interest were counted on consecutive CT images. The deWnition of the margin of the paranasal sinuses in CT datasets complicates the reproducibility. The boundary may easily be altered by various factors, such as mucosa variations, pathological diagnoses, by

Eur Arch Otorhinolaryngol (2009) 266:507–518

diVerent windows/CT value and by the resolution of the CT scanning unit. We included the mucosa in the segmentation area calculated to avoid variations. Table 3 demonstrates that all volumes measured by injecting materials (in italics) were smaller than using CT datasets (non-italics). Only the sphenoid volumes measured by Barghouth et al. [23] obtained smaller values due to the method used and their age range from 16 to 17 years. The comparison between female and male values revealed no signiWcant diVerences for the volumes of maxillary sinuses, as other authors reported [9, 23]. Our volumes obtained for the whole sphenoid and the right sphenoid sinuses are signiWcantly larger between genders. However, the left side of sphenoid sinuses does not diVer signiWcantly. This underscores the prevalence of an unbalanced septal partition between the right and left sides of sphenoid sinuses. Our total volumes of frontal and sphenoid sinuses were signiWcantly larger in men than in women. These results correspond to those of other research groups [24, 29].

Conclusion Our project develops automatic robot assistance in guiding the endoscope during FESS. For using CAS additional image-based information is needed. We have evaluated an approach to obtain the individual anatomy of the nasal cavity and the paranasal sinuses in manually segmented 3D models. Segmentation results can be used for robotic workspace deWnition, trajectory planning, surgical training or navigation during FESS. Our manually segmented model provides high accuracy but also requires a great deal of user interaction. Therefore, automatic segmentation is needed [19]. Depending on the complex anatomy, a model-based approach seems to be necessary. Our 3D models do provide high accuracy and the chosen CT datasets represent an average of the population for anatomical variations, extension and volumetric measurements. Therefore, our manually segmented 3D models can be used as a database for automatic modelbased segmentation. None of the manual and semi-automatic segmentation methods so far described provides risk segmentation. We consider the risk-orientated segmentation to be highly useful. Not being penetrated, the robot’s maximum distance to the segmented border can be adjusted according to the diVerently colored areas. Acknowledgments This work is part of the project “Robot-assisted intuitive endoscope navigation in endonasal surgeries with the help of preoperative computed tomography (CT) or magnetic resonance imaging (MRI) analysis” and we are grateful to the Deutsche Forschungsgemeinschaft (DFG) for funding this project. Bonfor, a research trust of the University of Bonn, is funding this project as well. The authors

517 wish to express their thanks to Prof. Dr. K. Schild and Priv.-Doz. Dr. med. Wilhelm of the Radiology Clinic of the University of Bonn for providing CT image data.

References 1. Strauss G, Fischer M, Meixensberger J, Falk V, Trantakis C, Winkler D et al (2005) Bestimmung der EYzienz von intraoperativer Technologie: WorkXow-Analyse am Beispiel der endoskopischen Nasennebenhöhlenchirurgie. HNO 54:528–535. doi:10.1007/ s000106-005-1345-8 2. Stammberger GH, Posawetz W (1990) Functional endoscopic sinus sugery. Concept, indications and results of the Messerklinger technique. Eur Arch Otorhinolaryngol 247:63–76. doi:10.1007/ BF00183169 3. Kennedy DW (2000) Functional endoscopic sinus surgery: concepts, surgical indications, and instrumentation. In: Kennedy DW, Zinreich J (eds) Diseases of the sinuses. Diagnoses and endoscopic management. Elsevier, Oxford, pp 197–210 4. Pérez-Piñas I, Sabaté J, Carmona A, Catalina-Herrera CJ, Jiménez-Castellanos J (2000) Anatomical variations in the human paranasal sinus region studied by CT. J Anat 197:221–227. doi:10.1017/S0021878299006500 5. Stallman JS, Lobo JN, Som PM (2004) The incidence of concha bullosa and its relationship to nasal septal deviation and paranasal sinus disease. AJNR Am J Neuroradiol 25:1613–1618 6. Oeken J, Bootz F (2004) Schwere Komplikationen nach endonasalen Nasennebenhöhlenoperationen. HNO 52:549–553. doi:10.1007/s00106-003-0861-7 7. Strauss G, Hofer M, Grunert R, Korb W, Trantakis C, Winkler D et al (2007) Ein Konzept für eine automatisierte Endoskopführung für die Nasennebenhöhlenchirurgie. HNO 55:177–184. doi:10.1007/s00106-006-1434-3 8. Yamashita J, Yamauchi Y, Mochimaru M, Fukui Y, Yokoyama K (1999) Real-time 3-D model- based navigation system for endoscopic paranasal sinus surgery. IEEE Trans Biomed Eng 46(1):107–116. doi:10.1109/10.736765 9. Kawarai Y, Fukushima K, Ogawa T, Nishizaki K, Gunduz M, Fujimoto M et al (1999) Volume quantiWcation of healthy paranasal cavity by three-dimensional CT imaging. Acta Otolaryngol Stockh 540:45–49. doi:10.1016/00016489950181198 10. Salah Z, Bartz D, Dammann F, Schwaderer E, Maassen M, Strasser W (2005) A fast and accurate approach for the segmentation of the paranasal sinus. Bildverarbeitung für die Medizin 2005. Informatik aktuel, Springer, pp 93–97. doi:10.1007/3-540-26431-0-20 11. Bolger WE, Butzin CA, Parsons DS (1991) Paranasal sinus bony anatomic variations and mucosal abnormalities: CT analysis for endoscopic sinus surgery. Laryngoscope 101:56–64. doi:10.1288/ 00005537-199101000-00010 12. Kantarci M, Karasen RM, Alper F, Onbas O, Okur A, Karaman A (2004) Remarkable anatomic variations in paranasal sinus region and their clinical importance. Eur J Radiol 50(3):296–302. doi:10.1016/j.ejrad.2003.08.012 13. Hiramatsu H, Tokashiki R, Suzuki M (2008) Usefulness of threedimensional computed tomography of the larynx for evaluation of unilateral vocal fold paralysis before and after treatment: technique and clinical applications. Eur Arch Otorhinolaryngol 265:725–730. doi:10.1007/s00405-007-0514-7 14. Wagner I, TingelhoV K, Westphal R, Kunkel ME, Wahl F, Bootz F, et al (2008) Ex vivo evaluation of force data and tissue elasticity for robot-assisted FESS. Eur Arch Otorhinolaryngol. doi:10.1007/ s00405-008-0644-6 15. Koulechov K, Strauss G, Dietz A, Strauss M, Hofer M, Lueth TC (2006) FESS control: realization of navigated control for

123

518

16.

17.

18.

19.

20.

21.

22.

Eur Arch Otorhinolaryngol (2009) 266:507–518 functional endoscopic sinus surgery. Comput Aided Surg 11(3):147–159. doi:10.1080/10929080600750789 Gill JD, Ladak HM, Steinman DA, Fenster A (2000) Accuracy and variability of a semiautomatic technique for segmentation of the carotid arteries from three-dimensional ultrasound images. Med Phys 27(6):1333–1342. doi:10.111811.599014 Asthon EA, Takahashi C, Berg MJ, Goodman A, Totterman S, Ekholm S (2003) Accuracy and reproducibility of manual and semiautomated quantiWcation of MS Lesions by MRI. J Magn Reson Imaging 17:300–308. doi:10.1002/jmri.10258 Apelt D, Preim B, Hahn HK, Strauss G (2004) Bildanalyse und Visualisierung für die Planung von Nasennebenhöhlen-Operationen. Bildverarbeitung für die Medizin. Informatik aktuell. Springer, Berlin, pp 194–198 TingelhoV K, Moral AI, Kunkel ME, Rilk M, Wagner I, Eichhorn KWG et al (2007) Comparison between manual and semi-automatic segmentation of nasal cavity and paranasal sinuses from CTimages. Conf Proc IEEE Eng Med Biol Soc 1:5505–5508. doi:10.1109/IEMBS.2007.4353592 TingelhoV K, Eichhorn KWG, Wagner I, Kunkel ME, Moral AI, Wahl FM, et al (2008) Analysis of manual segmentation in paranasal CT images. Eur Arch Otorhinolaryngol. doi:10.1007/ s00405-008-0594-z Lam WWM, Liang EY, Woo JKS, Van Hasselt A, Metreweli C (1996) The etiological role of concha bullosa in chronic sinusitis. Eur Radiol 6(4):550–552. doi:10.1007/BF00182491 Ariji Y, Ariji E, Yoshiura K, Kanda S (1996) Computed tomographic indices for maxillary sinus size in comparison with sinus

123

23.

24.

25.

26.

27.

28.

29.

volume. Dentomaxillofac Radiol 25(1):19–24. doi:10.1038/sj.dmfr.4600235 Barghouth G, Prior JO, Lepori D, Duvoisin B, Schnyder P, Gudinchet F (2002) Paranasal sinuses in children: size evaluation of maxillary, sphenoid, and frontal sinuses by magnetic resonance imaging and proposal of volume index percentile curves. Eur Radiol 12(6):1451–1458. doi:10.1007/s00330-001-1218-9 Spaeth J, Krügelstein U, SchlöndorV G (1997) The paranasal sinuses in CT-imaging: development from birth to age 25. Int J Pediatr Otorhinolaryngol 39:25–40. doi:10.1016/S0165-5876(96)01458-9 Uchida Y, Goto M, Katsuki T, Akiyoshi T (1998) A cadaveric study of maxillary sinus size as an aid in bone grafting of the maxillary sinus Xoor. J Oral Maxillofac Surg 56:1158–1163 Schumacher GH, Heyne HJ, Fanghänel R (1972) Anatomy of the human paranasal sinuses. 2. Volumetric measurement. Anat Anz 130(1):143–157 Anagnostopoulou S, Venieratos D, Spyropoulos N (1991) ClassiWcation of human maxillar sinuses according to their geometric features. Anat Anz 173(3):121–130 Karakas S, Kavakli A (2005) Morphometric examination of the paranasal sinuses and mastoid air cells using computed tomography. Ann Saudi Med 25(1):41–45 Emirzeoglu M, Sahin B, Bilgic S, Celebi M, Uzun A (2007) Volumetric evaluation of the paranasal sinuses in normal subjects using computer tomography images: a stereological study. Auris Nasus Larynx 34:191–195. doi:10.106/j.anl.2006.09.003