Monocular Surface Reconstructions in Transcanal Endoscopic Ear Surgery: A Phantom Feasibility Study

Abstract

To evaluate whether a computer vision model can utilize monocular endoscopic images of the middle ear to generate three-dimensional (3D) surface reconstructions that register to ground-truth (GT) imaging with clinically acceptable accuracy for surgical navigation.

Methods:

Eight temporal bone phantoms derived from the publicly available OpenEar data set were 3D-printed. Middle ear endoscopy was performed in each model using a standard monocular otologic endoscope, and a single endoscopic image captured of the mesotympanum. Images were processed with a monocular surface reconstruction model to generate predicted 3D surface reconstructions. Reconstructions were registered to GT meshes using natural anatomic landmarks. Surface reconstruction error (SRE) was quantified using the mean modified Hausdorff distance, and target registration error (TRE) was calculated using withheld test fiducials.

Results:

Mean SRE across models was 0.58 ± 0.10 mm, with 81.6% of points within 1 mm of GT surfaces. Mean TRE was 0.96 ± 0.17 mm.

Conclusion:

Simple 2D, monocular endoscopic images can be converted into high-fidelity 3D surface reconstructions of middle ear anatomy, which can be registered to ground-truth imaging with submillimeter registration accuracy. These findings establish the preclinical feasibility of monocular, endoscope-centric navigation for transcanal endoscopic ear surgery.

Introduction

The transcanal endoscopic approach to the lateral skull base (LSB) can be a useful alternative to conventional open approaches, as it avoids external incisions and brain manipulation^[1,2]. Despite these advantages, transcanal endoscopic LSB access remains high risk as critical neurovascular structures bound the surgical corridor. Surgical navigation has proven invaluable in skull base surgery by affording the localization of deep critical structures. Indeed, image-guided navigation in otology and neurotology was described more than 2 decades ago, with early work outlining frame-based and frameless stereotactic systems and emphasizing the need for accurate, minimally invasive platforms^[3].

Although no widely adopted navigation platform exists for transcanal LSB approaches, electromagnetic and optical tracking techniques have proven safe and reliable for skull base surgery generally. However, these systems require external hardware, patient-mounted reference frames, and explicit registration steps that introduce workflow complexity^[4,5]. A computer vision–based approach that derives registration directly from the live endoscopic image would reposition the endoscope itself as the sensing modality, reducing reliance on external tracking hardware and potentially simplifying intraoperative workflows.

Surface reconstruction, a computer vision technique that infers three-dimensional (3D) surface geometry from visual data, has been employed in surgery to generate surgical field maps that can be registered to segmented preoperative imaging^[6–13]. Although traditionally reliant on stereovision systems with paired video inputs, emerging models can now provide high-quality reconstructions from monocular/two-dimensional (2D) data. Advances in monocular surface reconstruction raise the possibility that imaging data from standard otologic endoscopes could support navigation-grade geometry for transcanal approaches. We hypothesize that a single mesotympanic image contains sufficient geometric detail to generate accurate 3D surface reconstructions that can be reliably registered to GT imaging–derived meshes with navigation-grade accuracy. This feasibility study represents a preclinical, phantom-based validation effort focused on quantifying spatial registration error of monocular surface reconstruction of the middle ear. In this preliminary work, we answer the fundamental question: “How accurately can a predicted 3D surface reconstruction of the mesotympanum be registered to the ground-truth middle ear surface?”

Methods

Data set source and 3D-printed phantom manufacturing

This phantom validation study utilized temporal bone models from the OpenEar library, a publicly available data set of 8 virtual 3D temporal bones. These models served as the GT meshes to which predicted surface reconstructions were registered and compared.

Three test fiducials were digitally embedded at the promontory and the anterior and posterior pillars of the round window niche to assess target registration accuracy (Fig. 1). Phantoms were fabricated on the Form 4 3D stereolithography printer.

Figure 1.

One of the 8 OpenEar models is pictured. Digital fiducials (yellow) are placed at the promontory, posterior and anterior pillars of the round window niche. Test fiducial points were placed at the apices of each pyramid.

Middle ear endoscopy and surface reconstruction processing

Following printing, middle ear endoscopy was performed in each phantom using a 0-degree, 2.7 mm diameter, 11 cm length Hopkins rod telescope coupled to a standard monocular endoscope (Karl Storz) and a single mesotympanic image, which captured the ossicular chain and the 3 test fiducials, was captured. Images were input into the surface reconstruction model, Visual Grounded Geometry Transformer (VGGT), to generate predicted 3D surface reconstructions^[14].

Natural landmark registration

Predicted surface reconstructions and their respective GT meshes were loaded into the biomedical imaging software, 3D Slicer. Natural landmarks were chosen as registration fiducials, as any clinical navigation system would require registration to be based on visible anatomy as opposed to embedded fiducials. Registration fiducials were placed at the following natural landmarks: malleus lateral process, umbo, and incus lenticular process in the GT mesh and the predicted surface reconstruction. Each predicted surface reconstruction was then registered to its GT mesh using these registration fiducials (Fig. 2).

Figure 2.

A, A single endoscopic image is processed with VGGT to generate a 3D surface reconstruction, which is coregistered to the GT mesh using predicted (red) and GT (green) natural fiducials. B, The surface reconstruction error (SRE) is then computed as the mean mHD between the registered reconstruction and the GT mesh. mHD indicates modified Hausdorff distance.

Surface reconstruction error

After registration, the surface reconstruction error (SRE, mm) was computed utilizing modified Hausdorff distances. The modified Hausdorff distance represents the minimum distance between each surface reconstruction point to its nearest surface in the GT mesh. These modified Hausdorff distances are then averaged to yield the mean Hausdorff distance (mHD). This metric reflects how well a predicted surface reconstruction approximates the GT surface^[15,16]. In addition, we report the percentage of predicted points that fall within 0.5, 1, and 2 mm of the GT surface. The natural fiducial registration and SRE workflow is depicted in Figure 2.

Target registration error

After registration, the apparent locations of the test fiducials in the predicted surface reconstruction were selected. The known locations of the test fiducials in the GT mesh were also selected. The target registration error (TRE, mm) was then computed as the Euclidean distance between each paired test fiducial. These TRE values are averaged to yield the mean TRE, which provides an assessment of how well a navigation system discretely matches known points. The workflow for TRE computation is demonstrated in Figure 3.

Figure 3.

A, Natural anatomic fiducials placed on the GT mesh and predicted surface reconstruction were used to perform coregistration. B, Independent test fiducials, not used in registration, were then identified on both GT and predicted surfaces. TRE was computed as the Euclidean distance between paired GT and predicted test fiducials. GT indicates ground-truth.

This study was deemed exempt from formal IRB review as it was not considered human subjects research.

Results

Surface reconstruction error

Key question: How close is the predicted surface to the ground-truth surface?

Across all 8 OpenEar phantom models, the predicted surface reconstructions demonstrated a close geometric match to their corresponding ground-truth (GT) surfaces. The mean surface reconstruction error (SRE) across models was 0.58 ± 0.1 mm (range: 0.47 to 0.7 mm). Across all reconstructed surface points, 50.5% were within 0.5 mm, 81.6% within 1 mm, and 99.2% within 2 mm of the ground-truth surface. Model-level SRE values are summarized in Table 1.

Table 1.

Summary of mean surface reconstruction and target registration errors per model

Model	Target Registration Error (TRE, mm)	Surface Reconstruction Error (SRE, mm)	Points Within 0.5 mm (%)	Points Within 1 mm (%)	Points Within 2 mm (%)
Alpha	1.1	0.48	60.0	85.5	100
Beta	1.1	0.65	51.9	77.6	96.6
Delta	0.95	0.54	62.5	88.4	100
Eta	1.1	0.5	58.5	93.1	100
Epsilon	0.84	0.7	44.4	79.3	98.6
Gamma	0.67	0.47	39.6	77.0	98.7
Theta	0.78	0.66	55.8	84.4	99.7
Zeta	1.1	0.67	31.5	64.7	100
Mean±STD	0.96 + 0.17	0.58±0.1	50.5±11	81.6±8.7	99.2±1.2

Target registration error

Key question: How well are discrete points localized after registration?

Target registration error (TRE) was computed using the 3 withheld test fiducials placed at the round window niche. The mean TRE across all models was 0.96 ± 0.17 mm (range: 0.67 to 1.13 mm). Individual model TREs are reported in Table 1.

Discussion

This work demonstrates navigation-grade surface reconstruction of the middle ear from a single monocular endoscopic image. In this 3D-printed phantom environment, we achieved an SRE of 0.58 mm and TRE of 0.95 mm, which are both below the accepted 1 mm registration error threshold for clinically useful navigation in skull base surgery^[17,18]. While the majority of reconstructed surface points fell within 1 mm of the ground-truth surface (81% on average), error increased at reconstruction boundaries. Nonetheless, our findings illustrate that accurate surface reconstruction of endoscopic data to ground-truth imaging—a fundamental step towards development of an endoscope-centric navigation system tailored to the transcanal corridor—is feasible.

Previous groups have achieved submillimeter/near-millimeter registration accuracy in the transcanal corridor. Schneider et al ^[19] combined surface matching of the external auditory canal, middle ear, and mastoid with optical tracking to localize deep targets, including the internal carotid artery, petrous apex, and jugular bulb, with near-millimeter accuracy. Uchida et al^[20] clinically implemented an augmented reality–assisted system using electromagnetic tracking for image registration, reporting submillimeter SRE but without evaluation of deep-target TRE. Our present work establishes the feasibility of surface reconstruction of the endoscopic video feed as a novel, alternative registration solution for guiding patient-to-CT registration in the transcanal corridor as opposed to surface mapping guided by an optical or electromagnetic tracker. Conceptually, a vision-based navigation system could perform surface reconstruction and registration passively during routine endoscopic inspection of the external auditory canal and middle ear, with optional reregistration as deeper dissection proceeds.

Vision-based navigation approaches, including SLAM, have been explored in skull base surgery. Bartholomew et al^[21,22] demonstrated stereo SLAM-based surface reconstruction with near submillimeter accuracy in the lateral and anterior skull base. Yet, their approach relied on stereo data, making it less practical for transcanal approaches where stereoscopic endoscopy is rarely performed. In contrast, the present work establishes feasibility with standard monocular endoscopy, which makes any subsequent navigation solution more accessible.

Our approach has key limitations. Validation was performed in uniformly colored 3D-printed phantoms, which do not fully replicate the complexity of live surgical environments. The impact of realistic color variation and occlusions (eg, bleeding) on reconstruction performance remains unknown and will require evaluation in cadaveric and in vivo settings. Unlike Schneider and colleagues, in this initial feasibility effort, we evaluated TRE using fiducials confined to the middle ear rather than deeper lateral skull base targets. Because registration landmarks and validation targets were spatially clustered along the ossicular chain, and TRE increases with distance from the centroid of registration points, accuracy at more remote structures cannot be inferred from these results^[23]. Achieving submillimeter TRE at deeper targets will likely require inclusion of deeper landmarks or staged reregistration as additional anatomic landmarks are exposed. In addition, the natural landmarks used for registration were readily accessible in phantoms but may require surgical exposure in vivo, potentially altering anatomy relative to preoperative imaging and affecting registration accuracy. Future validation must evaluate registration accuracy with the use of landmarks accessible through minimal surgical dissection. Finally, this work focuses on static surface reconstruction and registration; camera pose estimation, temporal stability, and maintenance of registration during endoscope motion were not assessed. These components are essential for clinical translation and represent the next phase of development for monocular, surface reconstruction–based navigation in the transcanal corridor.

Conclusion

This study demonstrates that a single monocular endoscopic image can generate a three-dimensional surface reconstruction that registers to ground-truth imaging with submillimeter accuracy. Future work will focus on cadaveric validation and integration of camera pose estimation and instrument tracking.