Automatic data-driven design and 3D printing of custom ocular prostheses

Loss of an eye due to trauma, painful blind eye, or untreatable tumours results in vision loss and visible facial difference. Following enucleation or evisceration, patients commonly wear a prosthetic eye over the conjunctiva to restore orbital volume and appearance. Conventional custom prostheses, often made from PMMA or glass, are hand-crafted by ocularists through labour-intensive, multi-step processes with variable outcomes. While additive manufacturing (AM) has gained traction in ophthalmology, existing digital methods still rely on manual CAD modelling, manual colour work, or impression moulds, and they often ignore either socket shape or appearance and have been validated in very small cohorts. This work addresses the need for a scalable, reproducible, and automated method that simultaneously captures socket shape and contralateral eye appearance to design and manufacture custom ocular prostheses.

Prior AM approaches for ocular prostheses have employed manual CAD based on CT scans of wax models, 3D scans of impression moulds or existing prostheses, manually set parametric models, average shapes, or manually segmented cone-beam CT. Some automated volume-difference reconstructions require CT, raising concerns for patients with cancer predisposition. Appearance reproduction has typically been omitted, hand-painted, or based on manually processed digital images without robust colour calibration; veining is often added manually. Reported studies generally involve very small samples (0–3 patients or cadavers) and do not describe a fully digital, automated, data-driven workflow that jointly addresses shape and appearance.

Digital end-to-end workflow: The patient’s anophthalmic socket and fellow eye are imaged with a modified Casia 2 AS-OCT device (1310 nm swept source, <6 mW). A set of 12 data-driven conformers (mirrored for laterality; total 24) is used to keep lids open, provide a shape prior, and define a circular window (~15 mm) for OCT. The socket scan comprises 256 slices capturing a 16 × 16 × 14 mm³ volume at 2145 × 1877 pixels per slice; the integrated colour camera (1280 × 1024, 12-bit, Bayer GR) captures the fellow eye under D50-matched illumination (high CRI 98, 45°/0° geometry). Data acquisition takes <30 minutes. Statistical shape model (SSM): 173 manually made prostheses were marked to standardize pose, scanned with a MEDIT T500, aligned, and parameterized via 838 corresponding vertices (closed surface mesh, 1672 faces). PCA via SVD yields an SSM with 17 principal modes explaining 98% variance. Reconstruction mean error is 0.27 ± 0.06 mm; leave-one-out generalization 0.31 ± 0.09 mm; specificity 1.10 ± 0.23 mm mean vertex distance. Conformer creation: 12 representative shapes (AAABBs between ~21.2×17.8×10.4 and 26.7×27.4×19.7 mm³) were converted into conformers by cutting a frontal window and carving a conical frustum; each was 3D printed (VeroClear) and polished. Shape extraction and fitting: OCT socket data are denoised and downsampled; column tracing detects anterior/posterior conformer window planes, then extracts the socket surface below the window. Optical path corrections (Snell’s law, refractive index 1.5; empirical −0.35d depth correction), alignment to iris plane, and a 6° gaze correction are applied. The visible socket area is ~143–174 mm² (mean 163 ± 9 mm²) versus a posterior prosthesis projection averaging 420 ± 73 mm² (range 230–736 mm²), implying about 61% of posterior surface information is missing and must be predicted. The prosthesis shape S(x) is fitted by minimizing E(x) = Edist + Eref using L-BFGS-B, with sequential mode enabling and bounds ±3 SD. Edist penalizes squared differences between the projected posterior surface of S(x) and the extracted socket depth map; Eref penalizes deviation from a target shape. If parameters hit ±3 SD, additional degrees of freedom (±30° rotation about y, ±2.5 mm translation along z) are introduced with regularization. Three shapes are predicted per patient using targets interpolating between the SSM mean and the selected conformer base (α = 0.0, 0.5, 0.9). Shape post-processing: The anterior surface is adjusted: the cornea is replaced by the mean shape, scaled to achieve a 2.5 mm apex and to match the iris diameter with smooth transitions. A locally varying 5% enlargement (excluding the cornea) and a clear coat thickness modulation are applied to ease clinical adjustment; curvature-based subdivision smoothing is performed. Regulatory safety excludes shapes with bounding boxes >30 × 29 × 20 mm³. Appearance synthesis: The fellow eye image is denoised, flat-fielded, colour-characterized end-to-end to CIELAB using a two-step camera characterization (RGB→XYZ linear transform; XYZ→LAB + root-polynomial correction) matched to D50/2° and 45°/0°. Specular highlights are detected in L* and inpainted. Iris: Limbus and pupil are detected using a modified multi-scale Daugman approach; the iris is unwrapped to a 4096 × 1024 texture; row-wise lightness contrast enhancement compensates print material light transport; pupil is set to pure black. Sclera: Watershed segmentation partitions pupil, iris, sclera, and aperture/skin; veins are removed from the sclera region (hue/chroma filtering). Nine sclera colours are extracted via k-means (k=9) to render staining using 3D Perlin-noise-based masks. A procedural, three-layer venous network is grown from anatomically motivated seeds using vein recipes controlling width, depth, branching, and path; ocularist-selected parameters are thickness th and branching ratio br. Fifteen vein profiles (sampled cross-sections with colour/transparency) are used to render veins with quadratic B-splines and depth-aware compositing. Digital model and fabrication: The OCT-provided iris mesh is refined (closed pupil, smooth limbus integration); a black inner cylinder behind the pupil ensures a dark pupil despite material translucency. Textures are mapped to the iris and sclera. Models (OBJ) are prepared with Cuttlefish using an ICC profile optimized for iris/sclera colours. Printing is performed on a Stratasys J750 (PolyJet, VeroVivid/Vero Clear/White/Black as applicable) via VoxelPrint/GrabCAD; reported print time is ~6 minutes per prosthesis for a full tray. Support is removed; parts are tumbled with ceramic chips and water to remove striations, then hand-polished and ultrasonically cleaned. Biocompatibility per ISO 10993 (analytical chemistry, in vitro, in vivo) confirmed toxicological safety. Clinical supply and adjustment: Ten patients at Moorfields Eye Hospital (MEH) were supplied between Nov 2022 and Apr 2023 under IRB approval (CA23/RE/960). At fitting, ocularists may remove clear material anteriorly to adjust lid closure or posterior fit material to tune iris position/size; typical adjustment session per prosthesis ~30–90 minutes (mean ~60).

• Automated end-to-end design and full-colour multi-material 3D printing of ocular prostheses was demonstrated in 10 clinic patients. • Labour reduction: Compared to current manual manufacture, the process requires approximately five times less ocularist labour while producing reproducible output. • Shape modelling: PCA-based SSM with 17 modes explains 98% variance; reconstruction mean vertex error 0.27 ± 0.06 mm; generalization 0.31 ± 0.09 mm; specificity 1.10 ± 0.23 mm. • Data capture: Socket visible surface extracted area averaged 163 ± 9 mm², whereas the posterior prosthesis projection averaged 420 ± 73 mm², implying about 61% of posterior surface information is predicted. • Fitting outcomes: For 8/10 patients at least one predicted shape could be adjusted and supplied; Patient 6 required no adjustments; Patients 1 and 3 (no orbital implants; unsuitable conformer/alignment issues) required scanning of an existing or modified prosthesis to supply. • Cosmesis assessment: Appearance- and colour-related scores (iris/pupil size/shape, iris and sclera colour and detail) were rated at least very good across patients. Patient satisfaction: 7 excellent, 3 very good. • Safety and production: ISO 10993 biocompatibility testing indicates toxicologically safe devices. Print time reported ~6 minutes per prosthesis for a full tray; digital model computation ~5 minutes; data acquisition <30 minutes.

The study demonstrates that a data-driven, automated pipeline can design and fabricate ocular prostheses that closely match socket geometry and contralateral eye appearance. By combining OCT-derived partial socket surfaces with a statistical shape model, the method effectively predicts the unseen regions and yields clinically adjustable shapes. Colour-characterized imaging with a printer-matched workflow reproduces iris size, structure, and colour as well as scleral staining and veining, which underpins high cosmesis ratings and patient satisfaction. Operationally, the workflow substantially reduces manual labour and standardizes output, enabling reproducibility and scalability, potentially allowing remote supply in underserved regions. The avoidance of alginate impressions and ionizing radiation (via AS-OCT) improves patient comfort and safety. The digital nature of the process supports continuous refinement based on new data and feedback, and the approach is being evaluated in a clinical trial for long-term performance against traditional methods.

This work introduces a fully digital, automated, and data-driven process for designing and 3D printing custom ocular prostheses that jointly optimize fit and appearance. Validated in 10 patients, the method achieves high cosmesis, reproducible outcomes, and markedly reduced labour compared to traditional craftsmanship. Future directions include expanding eligibility (e.g., complex sockets and pediatric cases) through enhanced OCT coverage and resolution, improved shape priors with larger datasets, better conformer design and alignment robustness, and advances in printer materials and colour management. The approach may generalize to other prostheses (dental, facial) and hybrid solutions (e.g., cosmetic covers) within a scalable digital manufacturing ecosystem.

Most cases still require final clinical adjustments by an ocularist, making the devices modified custom prostheses. Very complex sockets may not be captured adequately by OCT or represented by the SSM, necessitating traditional craftsmanship. Approximately 80% of patients are currently eligible; patients with conditions such as nystagmus or strabismus may not provide usable scan data. Outcomes depend on correct conformer selection and scan alignment; soft tissue displacement (notably in patients without orbital implants) can degrade shape prediction. Safety-imposed bounding box limits can exclude deeper shapes. Economic efficiency requires sufficient case volume. Pupil size matching is affected by ambient lighting; scleral colour/veining can vary with eye irritation.