Introduction

Recent studies suggest that following a sports-related anterior cruciate ligament (ACL) tear, anatomic graft placement results in superior ACL reconstruction outcomes with respect to knee function and return to sport in young patients [11, 18, 27]. Magnetic resonance imaging (MRI) of the knee readily shows the cruciate ligaments and their attachments to bone in multiple planes [4, 7, 13] and also demonstrates remnant tissue at attachments even after an ACL tear [11]. If patient-specific ACL attachment locations can be reliably identified on pre-operative MRI, surgeons could potentially use this information intra-operatively to ensure ACL reconstruction tunnels are placed in the appropriate anatomic locations.

Clinical knee MRI traditionally uses a set of two-dimensional (2D) sequences optimised to diagnose injury to ligaments, menisci, bone and cartilage [17, 20]. However, spatial resolution is limited by the relatively thick slices (3–4 mm) on 2D sequences. Use of a 3D MRI base sequence with smaller isotropic voxels [15] might improve intra- and inter-observer reliability in localising ACL attachments. Localising ACL attachments on pre-operative MRI for anatomic guidance is an emerging concept, and we could not find any published studies comparing 3D and 2D sequences for this application.

The purpose of this study was to test whether the intra- and inter-observer reliability of localising ACL attachments on MRI was significantly improved when using a high-resolution 3D base sequence compared with routine clinical multi-planar 2D base sequences, in a heterogeneous paediatric patient group with both ACL deficient- and intact knees. Paediatric patients were studied since the need for precise tunnel placement is arguably greatest in the youngest patients, whose physes may still be open and whose grafts must function for a long lifespan. It was hypothesised that that a high-resolution 3D sequence would permit identification of ACL attachment locations with higher intra- and inter-observer reliability than routine 2D MRI.

Materials and methods

Ethical approval for this retrospective study was received from the University of Alberta research ethics panel (study 0030905). As part of protocol development, our group has been intermittently adding a 3D multiple-echo data image combination (MEDIC) gradient echo (GRE) series to our routine knee MRI protocol for 5 years. We reviewed all paediatric patients that underwent a knee MRI between July 2005 and October 2013 to assess an acute sports injury at two tertiary hospitals and a private radiology clinic. From these, we selected 23 consecutive paediatric patients whose knees were scanned with both 2D proton density (PD)-weighted sequences (slice thickness 3–4 mm), and 3D MEDIC GRE (slice thickness 0.8 mm). After applying exclusion criteria of prior knee surgery, known pre-existing ligament injury, associated avulsion fractures or PCL tears, we included 20 knees. Of these, ten knees had intact ACL and ten knees had a high grade or full thickness ACL tear visible on MRI and confirmed surgically. Included patients were aged 10–17 years at time of MRI, all had intact PCL, 11/20 were female, and average age was 14.9 ± 2.6 year (mean ± SD).

Imaging

Knee MRI was performed using standard clinical protocols (Siemens AG, Munich), on Siemens 1.5-T scanners. 2D imaging was in three planes with slice thickness of 3–4 mm. Scans were co-localised (i.e. coordinates on the same plane from same localiser). 2D sequences included coronal and sagittal PD-weighted images (typical parameters: thickness 3.0 mm, 512 × 256 matrix, 16 × 16 cm FOV, TR/TE 1,800/13 ms) and axial PD weighted with fat saturation (PD FS; thickness 4.0 mm, 384 × 230 matrix, 16 × 16 cm FOV, TR/TE 3,000/42 ms) images through the knee. GRE is commonly used for clinical 3D sequences. Therefore, 3D imaging consisted of a 3D T2-weighted GRE sequence (MEDIC GRE, thickness 0.8 mm, 256 × 248 matrix, 20 × 20 cm FOV, TR/TE 33/19 ms).

Image processing

All image analysis was conducted off-line using custom software in the MATLAB programming environment (MATLAB 7.8.0; The MathWorks, Natick, USA). To localise ligament attachment sites on MRI, points along the ACL and PCL attachments were manually identified on all slices in which they were clearly visible. Users were instructed to employ a conservative approach to ligament point selection and to avoid the extreme edges of ligaments, which were more likely to be affected by volume averaging. Observers identified ligament attachments as points where linear structures of low intensity in PD images or intermediate intensity in MEDIC GRE images directly contacted bone in the expected anatomic location (Fig. 1). For patients with torn ligaments, observers identified remnant tissue present at bony attachments. A model of the bone surfaces was also generated for 3D visualisation of ligament attachment points (Fig. 2). Note that posterior cruciate ligament (PCL) attachments were identified as an internal control, since the PCL is a well-defined, thick and dark (hypointense) structure on PD-weighted MR and consequently more easily delineated than the ACL in our experience.

Fig. 1
figure 1

Identification of the ACL femoral attachment sites on 2D PD slices (a, b) and on 3D MEDIC GRE (d, e) on knees with intact ACLs. After the user-identified points at all visible ACL attachment sites on all sequences, the cluster of these points was projected onto a best-fit plane, and the area enclosed by the attachment sites and the centre of the enclosed area were calculated and visualised (c, f). On plots (c, f), the open circles (arrowhead in c) represent attachment points as found on the surface of the bone and solid black circles (straight arrow in c) represent each point projected onto the best-fit plane. The centre of the enclosed area (curved arrow in c) is represented by the larger circle

Full size image
Fig. 2
figure 2

Reconstruction models of bone surfaces showing the 3D locations of the ACL femoral attachments (dark dots) in knees with intact ACLs, generated from a 2D PD and b 3D MEDIC GRE sequences. Note that the model in (a) has flat femoral condyle tips due to volume averaging

Full size image

The cluster of user-identified points representing each ligament attachment was used to calculate the centre and area of the region bounded by these points on a best-fit plane for each ligament (Fig. 1), for both 2D and 3D sequences. The 3D distance between the centres of ligament attachments determined separately by two observers was recorded as D 12. To compare the attachment areas selected by two observers, points from both users were projected onto a single combined best-fit plane and the percentage of area overlap (Ov %) was calculated, as the area of overlap divided by the average of the two areas.

Reliability analysis

ACL and PCL bony attachments were identified on the 20 knees by two observers (User 1 and User 2) on 3D MEDIC GRE sequences and 2D PD sequences. Inter-observer reliability (comparison of ligament attachment centres and areas identified by User 1 vs. User 2) and intra-observer reliability (i.e., test–retest reliability for the ‘test’ of identifying ligament attachments, by comparison of ligament attachment centres and areas identified by User 2 on two separate occasions) were calculated. To assess intra-observer reliability, User 2 identified ligament attachments on a second occasion with a 2-week interval between reading sessions, blinded to original data. User 1 was a radiologist with dual musculoskeletal and paediatric fellowship training, and User 2 was a trainee with experience in knee MR anatomy. Users were blinded to clinical and demographic data and to the location of ligament attachments identified by the other user or by themselves at a previous session. Both users recorded the time required to identify ACL and PCL attachments on MR images.

Statistical analysis

Intra- and inter-observer reliability of identification of ACL and PCL attachments was tested in several ways. Descriptive statistics were generated for the 3D distance (D 12) between the centres of the attachments generated by the two observers (reported at an accuracy of 0.1 mm) and for the percentage overlap of ligament attachment areas between observers (reported to the nearest whole percentage). Intra-class correlation coefficients (ICCs) for ligament attachment areas were calculated for single measures using a two-way random effects model. Statistics were calculated on SPSS (Chicago, IL, USA, v.19). Continuous variables were described by mean ± SD when appropriate. Using a level of significance of p = 0.05, we tested differences between means of continuous variables such as D 12, attachment areas, and Ov % by unpaired, two-tailed Student’s t tests after confirming approximately normal distributions from review of frequency tables. For calculation of sample size, we noted that the inter-observer mean difference between ligament centre 3D locations was at most 2 mm (half the axial 2D slice thickness), with SD = 1 mm, when identified using 2D sequences [31]. To detect improvement of the inter-observer mean difference to 0.5 mm (approximately half the 3D GRE slice thickness), at α = 0.05, β = 0.80, the minimum required n = 7. The sample size exceeded this requirement. Post hoc analyses of statistical power were also conducted to assess non-significant comparisons.

Results

Ligament attachment centres

Table 1 shows the distances D 12 (distance between attachment centres, for inter-observer or intra-observer comparisons). In all cases, the mean inter-observer difference was <2.0 mm with a 95 % confidence interval (CI) <4.0 mm. There were no significant differences in D 12 whether the 3D or routine 2D MRI sequence was used, regardless of ACL tear status (n.s.). The range of intra-observer variability was less than or roughly equal to inter-observer variability in all cases.

Table 1 Inter- and intra-observer reliability of ACL and PCL attachment site centres
Full size table

Ligament attachment areas

Table 2 shows the inter-observer comparison of ACL attachment areas identified by each user and associated measures of agreement.

Table 2 Inter-observer reliability of ACL and PCL attachment site areas
Full size table

For the inter-observer comparison, the ligament attachment areas identified by User 2 were lower than User 1, with absolute differences of 14–38 mm2 for 3D sequences and 19–25 mm2 for 2D sequences. This difference reached statistical significance for the ACL femoral attachments in 3D sequences. The overlap in ACL attachment areas between User 1 and User 2 ranged from 72 ± 11 to 80 ± 9 %. There was no significant difference between inter-observer area overlap in knees scanned with 3D sequences versus 2D sequences. Post hoc analysis showed we had 89 % power to detect a 15 % difference in area overlap. Interestingly, for the ACL attachments, there was a non-significant trend towards improved inter-observer Ov % with 2D sequences over 3D sequences. Intra-observer analysis of attachment areas showed equivalent or better agreement than inter-observer analysis, with similar trends (data not shown).

The time required to mark ACL and PCL attachment points for the inter-observer analysis was significantly lower on 2D PD sequences versus 3D MEDIC GRE sequences (5.7 ± 1.4 min vs. 9.3 ± 1.7 min for User 1, p = 0.002; 8.9 ± 1.9 min vs. 11.9 ± 2.9 min for User 2, p = 0.008).

Discussion

The most important finding of this study was that there were no significant differences in inter- or intra-observer reliability of identifying knee ACL attachment centre locations (D 12) whether a high-resolution 3D or routine 2D MRI sequence was used. This was true despite the substantially higher out-of-plane spatial resolution of 3D MEDIC GRE sequences over routine multi-planar 2D PD (slice thickness 0.8 mm vs. 3.0–4.0 mm). Users identified centres of ACL bony attachments with mean inter-observer error <2.0 mm and 95 % confidence interval (CI) <4.0 mm (Table 1) in all cases regardless of ACL tear status. This degree of variation is small relative to attachment size and clinically acceptable.

The mean error <2.0 mm is of the same magnitude as the 2 mm mean difference in ACL footprint dimensions between 3D MRI and cadaveric dissection reported by Han et al. [8] using dedicated high-resolution 0.6 mm slice thickness 3D MRI. The <2.0 mm mean error is also small relative to ACL attachment cross-sectional dimensions, reported at 14–29 × 8–11 mm at tibial attachment, and 14–23 × 7–11 mm at femoral attachment [3, 5, 9, 15, 21, 24, 29]. As well, the mean error is smaller than the 3 mm diameter of a typical Steadman bone awl used to place the centre of the graft tunnel [25, 32].

Several 3D GRE sequences, including MEDIC [1, 16, 28], have previously been used to diagnose knee ligament and cartilage abnormalities [2, 6, 10, 22, 34]. However, users noted that MEDIC GRE sequences provided relatively poor image contrast at ligament edges, as others have observed [14]. Though 3D MEDIC GRE had improved spatial resolution compared with 2D PD sequences, this was likely offset by a loss of contrast resolution at the ligament-bone interface, resulting in no net gain in inter-observer reliability (Fig. 3). Using current scanners, it is impractical to obtain high-resolution 3D data with thin slices on PD sequences for clinical use, due to lengthy acquisition times [8].

Fig. 3
figure 3

PCL femoral attachment visualised on a sagittal 2D PD slice (a) and a 3D MEDIC GRE slice (b). Compared with 2D PD sequences, 3D MEDIC GRE gives poorer contrast resolution at the ligament-bone interface found near ligament edges (see arrows). Also note several other features of the 3D GRE image compared with 2D PD: higher contrast between articular fluid (bright) and bone (dark) on GRE than PD; the generally coarser appearance of the GRE image (b), due to lower in-plane resolution, and the darker and more heterogeneous marrow appearance in (b), which is likely due to gradient blooming at trabeculae

Full size image

Inter-observer reliability of attachment areas was higher for 2D sequences, with lower ICCs when using the 3D sequence. This may be because attachment edges are more confidently depicted on PD sequences than on the 3D GRE sequences. However, the difference in attachment areas between observers (User 1 vs. User 2) was substantial when using either 2D or 3D sequences. This suggests that additional consensus training would be beneficial. ACL attachment centres were localised with substantially higher reliability than attachment areas and represent a more clinically useful landmark for anatomic tunnel placement than attachment areas.

The results of this study are of interest to surgeons involved in management of ACL-injured young patients. One justification for pre-operative knee MRI imaging in ACL tears is for potential use as a tool for intra-operative guidance of anatomic ACL reconstructive surgery [8, 11, 19, 26]. Anatomic tunnel placement can be challenging even for experienced surgeons. We have shown that ACL attachment centres can be identified with high inter- and intra-observer reliability on both 2D and 3D base MRI sequences, regardless of ACL tear status. As a potential application of these findings, a patient-specific 3D knee model showing ACL attachment centres (Fig. 2) could be produced and rotated to mimic the arthroscopic portal view seen intra-operatively (Fig. 4). Surgeons could use this view and measurements made from the 3D model to assist in placing ACL graft tunnels in anatomic locations, possibly by identifying the attachment centre in relation to arthroscopic landmarks such as the posterior capsular insertion or inferior articular margin. Initial validation of this potential application is currently underway.

Fig. 4
figure 4

Schematic intra-operative visualisation of ACL attachments on a 3D model. The right-sided monitor shows a 3D knee reconstruction of the patient’s femur and tibia, with ligament attachments (red) and estimated centres (green) marked, created by custom software after users identified ACL attachments on MRI slices. The left-sided monitor shows the medial portal arthroscopic view, with visible landmarks including the lateral femoral condyle cartilage margin and posterior bony margin. Prior to surgery, the 3D model would be rotated into the viewing angle of a standard medial arthroscopic portal applied. The 3D model could be used both pre-operatively (to measure distances to landmarks such as the posterior capsular margin and inferior articular margin) and intra-operatively (to directly visually aid in locating the centre of the graft tunnel to be placed). Note that this image is conceptual only, and the image on the right monitor has been added digitally

Full size image

The results of this study are also of interest to clinicians who order knee MRI and to those who plan knee MRI protocols. It has been shown in this study that adding a high-resolution 3D sequence did not substantially improve reliability of ACL attachment localisation over simply combining information from existing routine 2D sequences. Surgeons in clinical practice need not order a lengthy dedicated 3D MRI to localise ligament attachments, but can confidently use a standard routine 2D MRI, such as might be submitted digitally from an outside institution or performed unsupervised on nearly any MR scanner, for this application. Routine 2D knee MRI as already performed in many centres, including many smaller peripheral centres, fortunately provides data sufficient to reliably model ACL attachment locations in 3D. With this result, once the MRI-guided ligament attachment technique is validated, it could be applied broadly in real-world practice.

This study was performed on paediatric knees. The need for precise tunnel placement is arguably greatest in the youngest patients, whose physes may still be open and whose grafts must function for a long lifespan. Paediatric knees also generally lack degenerative changes, which might decrease the reliability of our technique in model development. However, the findings of our study should be equally applicable to adults without substantial degenerative changes. Also, a prior study showed that ACL tear status did not influence the reliability of localising ACL attachment centres on MRI, and so, we included patients with both torn and intact ACLs in this study [31].

This study has limitations. First, this retrospective study was designed to assess reliability, not validity, of localising ACL attachment centres. The 3D attachments seen at MRI have face validity, lying visually at the expected locations of the native ACL seen arthroscopically and anatomically [31]. Quantifying this accuracy requires prospective surgical/MRI study in patients and/or cadavers. Second, the 2D sequences in this study have been refined over years of clinical use, while the specific 3D sequence used has seen less clinical testing in the knee. Also, other 3D sequences exist. In particular, several groups have recently used 3D sampling perfection with application-optimised contrasts using different flip-angle evolutions (SPACE) MRI sequences to assess cruciate ligaments and menisci [12, 23, 30, 33]. However, SPACE sequences have T2-weighting, which like the 3D MEDIC GRE used in our study is likely not optimal for image contrast at ligaments. It is possible that a 3D sequence could be further optimised to improve on the reliability demonstrated here.

Conclusions

ACL attachment locations were not identified more reliably using a high-resolution 3D MEDIC GRE MR sequence than by combining multi-planar routine 2D PD knee MR sequences, regardless of ACL tear status. All models had an average distance of <2.0 mm between user-identified ligament attachment centres. Based on this study, it is not necessary to add a 3D MEDIC MRI sequence to routine knee MR protocols for purposes of reliable ACL attachment localisation. Prospective surgical study in cadavers and patients is warranted to further validate the clinical application of the promising radiologic findings demonstrated in this study.