Orbital Volume Segmentation AI (Human-in-the-Loop U-Net)
Automatic bony-orbit segmentation for orbital volume measurement
▸ More attributes
Brief overview
This study developed a U-Net–based deep learning platform with multiple pretrained CNN encoders (including MobileNet-V2, ResNet34/50, InceptionResNet-V2, and EfficientNet-B0) to segment orbital volume from 3D facial CT scans. Across 349 retrospectively collected CTs from a single Korean institution, the final MobileNet-V2 U-Net achieved Dice coefficients up to 94.1% with inference times of ~33 ms per axial slice. The authors also evaluated a human-in-the-loop semi-automated workflow, showing faster annotation times than manual segmentation without loss of accuracy. No external validation, prospective testing, public code release, or clinical deployment was reported. The model was designed as a research-stage decision-support tool for orbital surgical planning and volumetric assessment.
Manual-only dataset (178 CTs): Dice ≈ 0.90.
Manual-only dataset (178 CTs): Dice ≈ 0.90. Full dataset with AI-assisted refinement (349 CTs): best Dice = 0.94 (MobileNetV2 U-Net). Reader study: AI-assisted segmentation improved Dice (90.3% vs 88.7%) and reduced annotation time by 22–80%.
| Study design | Retrospective cohort |
|---|---|
| Center type | Single-center |
| Patients | 349 |
| Images / eyes | 349 CT images |
| Split | ~60/20/20 |
| Reporting frameworks | None stated |
| Age range | >19 |
|---|---|
| Sex (% female) | Not reported |
| Race / ethnicity reported | N |
| Subgroup performance | N |
- Concept
- Prototype
- Internal val.
- External val.
- Regulator
- Clinical
▸ Why this score?
The APPRAISE-AI tool grades AI/ML medical studies across 24 items in 6 subdomains, scored independently by two reviewers. The total reflects the methodological transparency of the publication — not the underlying clinical value of the model.
How to read this: high scores in clinical relevance and reporting with low scores in methodological conduct, robustness, or reproducibility is the dominant pattern in ophthalmic AI today.
Quality bands: Low <40 · Moderate 40–59 · High 60–79 · Very High 80+. Most published ophthalmic AI sits in the Low–Moderate band; studies that break into High typically combine multi-center datasets, prospective evaluation, and open-source release.
Bibliographic record ▸
| Model ID | oculoplastics_chang_2025 |
|---|---|
| Full author list | Cho, J., Shon, B., Choi, K. Y., Jeong, S., & Ryu, J. Y. |
| Journal | J Craniomaxillofac Surg |
| DOI | 10.1016/j.jcms.2025.01.007 |
| Conditions | Orbital volume assessment |
Model internals ▸
| Architecture family | CNN |
|---|---|
| Architecture detail | U-Net with ImageNet-pretrained encoders, paired with a human-in-the-loop workflow (AI pre-segmentation + clinician refinement). |
Disclosures ▸
| Funding | MSIT(Ministry of Science and ICT), Korea, under the ITRC(Information Technology Research Center) support program(IITP-2025-RS-2020-II201808) supervised by the IITP(Institute of Information & Communications Technology Planning & Evaluation); Korea Health Technology R&D Project grant (grant number: HR22C1832) |
|---|---|
| Conflicts of interest | NR |
Status, version, reviewers ▸
| Card status | Published |
|---|---|
| Card version | v1.0.0 |
| First added | 2026-04-28 |
| Last reviewed | 2026-04-28 |
| Reviewers | attribution opt-in pending |
Card changelog ▸
- 2026-04-28 initial publication (v1.0.0)