The overall workflow is summarised in Fig. 1. Written informed consent was taken from patients recruited to this ethically approved preliminary study. The HoloLens software was used only within the institution where it was developed. Preoperative contrast-enhanced CTA scans were performed using a 256-slice Philips Brilliance CT scanner (Koninklijke Philips N.V., Amsterdam, The Netherlands). The scans were undertaken with patients in the prone position to limit tissue deformation of the soft tissues of the calf and to reduce compression of the perforating veins. In certain cases, this was not possible due to other injuries limiting prone positioning of the patients. As an alternative, the respective legs were elevated to prevent compression of the calf skin and muscle. This was done to mitigate the effect of anatomical deformation between acquisition time and intervention and thereby minimise registration error. In order to distend the lower leg perforators, a tourniquet was applied above the knee level. To obtain both arterial and venous phases within a single contrast scan, a phased administration of contrast agent was performed. Specifically, a split bolus comprising two 70-mL volumes of Omnipaque (General Electric Healthcare, Chicago, IL, USA) was administered intravenously, followed by a saline bolus chaser. Typically, the images were acquired with an axial in-plane resolution of 0.7 mm and slice thickness of 0.9 mm.
The CTA volumes were segmented using the Vitrea 6.7.4 software (Vital Images, Inc., Minnetonka, MN, USA) into skin, bone, muscle, and vascular models by the consultant radiologist. The vascular models were produced with segmentations of both venous and arterial vessels of the lower leg. Where appropriate, the trunks of the medial sural artery perforators (MSAPs) were segmented together with the associated site of perforation through the muscle fascia. The lengths of the trunk and its branches were calculated using the Vitrea built-in vascular tools and the approximate site of each perforator was measured relative to the medial femoral epicondyle. Skin segmentations were performed using the Vitrea autoskin function, whereby a combination of voxel thresholding and morphological operations rapidly identifies the outer tissue layer. This facilitates accurate hands-free registration of models to their respective patients. In addition, the bone segmentations allowed for accurate correlation with anatomical landmarks such as the tibial tuberosity. Before surgery, each case was discussed with the surgical team to explain the models and anatomy and to confirm the choice of perforator. It was not uncommon to mark a selection of perforators so that a decision could be made at the time of surgery.
The segmented volumes were then loaded as Digital Imaging and Communications in Medicine (DICOM) files into ITK-SNAP 3.6.0 [15], where minor refinements were made through meticulous use of its region growing and manual painting functionality, and where mesh representations were generated by the marching cubes algorithm [16]. These were further manipulated with MeshLab 1.3.3 [17] for smoothing and mesh complexity. Respectively, the manipulations performed were Humphrey’s Classes Laplacian smoothing [18], followed by edge collapse decimation using a quadric error metric [19]. These resulted in optimised anatomical representations, increasing the performance of the HoloLens application while minimising loss of model precision. Tests were performed to check whether the loss of precision was within the clinically acceptable range (5 mm) for this specific application. For example, the axial length of the skin model for Case #1 was reduced by 0.92 mm over a total of 591.6 mm, equal to a change of 0.16%.
Written within the Unity framework, version 2017.1 (Unity Technologies, San Francisco, CA, USA), a custom-developed HoloLens C# Universal Windows Platform application was utilised at the time of patient marking. Once launched, the application generated ‘holographic’ overlays, correctly rendered for both left and right eyes, at a default distance and rotation with respect to the wearer’s coordinate frame. Subsequently, using a combination of spatial translation and rotation hand gestures, the operating surgeon manipulated the virtual anatomy from a static posture until a satisfactory degree of anatomical landmark and skin outline alignment against the anaesthetised patient were achieved.
Through either voice commands or a toolbar button, the user interface permits switching between translation and rotation modes. In both cases the HoloLens ‘air-tap and hold’ gesture was used as a source of 3D motion input. Having been employed successfully in other image guidance applications [20], the ‘rolling ball’ control mechanism [21] was adopted as it provides a very intuitive way of manipulating orientation that is independent of observer viewpoint. The spatial scale information embedded in each CTA scan was retained throughout the segmentation and object building process, so that no scale adjustments were required following HoloLens model import. The required procedural changes were minimal. Having annotated the skin with a sterile marker pen under HoloLens guidance, the target positions were compared to the sites of the perforator vessels as subsequently identified by audible Doppler ultrasound and surgical appearances.