An approach to automatic evaluate registration performance for image guided navigation

Purpose

Three dimensional image guided navigation is increasingly used in operating room (OR) for various surgical cases, especially for brain surgeries, and greatly facilitate the procedures. The registration between image space and physical space is crucial for this technology. Many registration algorithms and methods have been proposed and implemented into image guided navigation devices that claim certain accuracy, which indicates how accurate the surgeons can position a target inside of the patient in average. For particular case, the registration error could be off from this average number due to human error or other unexpected factors. However, a reliable “indicator” for the registration accuracy lacks through the surgical procedures. Several commercial image guided navigation devices based on fiducial marks provided fiducial registration error, which is uncorrelated to the target registration error and proven a poor predictor of registration accuracy [1]. Other devices usually require surgeons to visually and qualitatively estimate the registration accuracy by checking the corresponding on landmarks and surfaces between pre-op scan images and patients [2].

Intra-operative mobile cone beam CT devices such as xCAT® (Xoran Technologies, Inc., Ann Arbor, MI, USA) have emerged in recent years as a promising facilitator of minimally invasive image guided surgery (IGS) [3]. Not only did this device provide surgeons with near real-time image updates, but also with the coordination of patient in the OR. With this kind of device, surgeons are allowed to conduct image based on registration between image space (pre-operation scan image volumes) and patient space (intra-operation scan image volumes), without troublesome interactions of manual probing (locating fiducials or landmarks on patients, or swiping the surface of a desired area of patient, with tracked probe). Case-specific quantitative registration accuracy information can be very helpful for surgeons to understand how accurate they can rely on the navigation tools for this particular case, or if they need to redo the registration. Automatic segmentations on both pre-op and intra-op image volumes provide surface geometric information of patient face; after image based-on registration, geometry matching of such surfaces from both image volumes is used for evaluating registration accuracy. Phantom experiments show that this method is intuitive and easy to implement, and clinical application is under investigation through our device development project.

Methods

Phantom image acquisitions in hospital and with xCAT: A head phantom mounted with 12 disposable skin markers (Beekley CT-SPOTS®) on its outside surface and in a hollow chamber was scanned by both a traditional hospital CT system and with Xoran Technologies’ xCAT intraoperative CT system. For the hospital CT system, the image sampling was 0.486 × 0.486 × 1.00 mm while for the xCAT it was 0.4 × 0.4 × 1.2 mm. The xCAT scan was performed with a low-dose protocol (0.4 mSv) which is approximately one seventh the x-ray dose of traditional head CT. A CT and MR image from a patient was also acquired, in both cases using traditional hospital scanners. The CT and MRI image volume resolutions were 0.654 × 0.654 × 4.0 mm and 1.2 × 1.2 × 4.0 mm, respectively.

We calculated the registration error by segmenting the frontal surface of the head in the different image volumes and analyzing their mismatching. Since the skin can easily be identified in both MR and CT images, the geometric information of frontal surface was generated through automatic image processing and segmentation. For pre-op CT scan case, we were also able to calculate the registration error through skin markers since they were visible in both pre-op and intra-op image volumes, to show consistency between these two evaluation methods. Using image intensity thresholding, the markers were automatically segmented and the center of each marker was calculated both in the xCAT image and in the pre-op scan images. The registration error was estimated by calculating the average location differences of all markers (target registration error). For pre-op MR case, we only calculated the registration error by analyzing the frontal surface mismatching. Further evaluation might be helpful when multi-modality skin markers are available for pre-op MR scan.

Results

The calculated registration error based on the frontal region surface matching was 0.18 ± 0.20 mm (mean ± SD) for pre-op CT scan case. The matching errors of skin markers are given in Table 1. The estimated registration error based on the surface matching is consistent to the registration error level based on target markers. The average error of these markers was 0.24 mm, with standard deviation 0.16mm.

Table 1.

Error estimation based on skin markers

marker	markers in xCAT volume			markers in hospital CT volume			error
number	x	y	z	x	y	z
1	284.42	251.85	97.66	284.55	251.94	97.53	0.18
2	354.36	241.22	68.07	354.45	240.91	67.96	0.19
3	379.03	285.98	96.52	379.14	285.86	96.44	0.11
4	279.93	244.36	67.91	279.88	244.08	67.80	0.17
5	372.71	327.32	77.42	372.92	327.08	77.27	0.22
6	260.95	319.81	79.33	260.92	319.80	79.41	0.09
7	323.10	210.78	83.35	323.26	210.42	83.21	0.23
8	464.87	147.82	61.23	465.07	147.25	60.96	0.40
9	497.09	266.31	27.63	497.07	265.49	27.15	0.66
10	123.98	283.87	47.23	123.82	283.81	47.19	0.08
11	160.95	151.99	53.04	160.74	151.91	52.90	0.19
12	315.04	70.42	44.43	314.91	69.97	44.14	0.39
					Average error		0.24
					Standard Deviation		0.16

For pre-op MR scan, we calculated registration error as 0.80 ± 0.52mm.

Conclusion

Analyzing the mismatching between segmented frontal surfaces of head from pre-op and intra-op image volumes could be a practical approach to estimate the registration error through surgery, which can provide surgeons case specific and quantitative accuracy information for surgical navigation. No user interaction is needed to generate this useful information through the surgery.