Knoedler, Leonard and Alfertshofer, Michael and Simon, Siddharth and Prantl, Lukas and Kehrer, Andreas and Hoch, Cosima C. and Knoedler, Samuel and Lamby, Philipp (2023) Diagnosing lagophthalmos using artificial intelligence. SCIENTIFIC REPORTS, 13 (1): 21657. ISSN 2045-2322,
Full text not available from this repository. (Request a copy)Abstract
Lagophthalmos is the incomplete closure of the eyelids posing the risk of corneal ulceration and blindness. Lagophthalmos is a common symptom of various pathologies. We aimed to program a convolutional neural network to automatize lagophthalmos diagnosis. From June 2019 to May 2021, prospective data acquisition was performed on 30 patients seen at the Department of Plastic, Hand, and Reconstructive Surgery at the University Hospital Regensburg, Germany (IRB reference number: 20-2081-101). In addition, comparative data were gathered from 10 healthy patients as the control group. The training set comprised 826 images, while the validation and testing sets consisted of 91 patient images each. Validation accuracy was 97.8% over the span of 64 epochs. The model was trained for 17.3 min. For training and validation, an average loss of 0.304 and 0.358 and a final loss of 0.276 and 0.157 were noted. The testing accuracy was observed to be 93.41% with a loss of 0.221. This study proposes a novel application for rapid and reliable lagophthalmos diagnosis. Our CNN-based approach combines effective anti-overfitting strategies, short training times, and high accuracy levels. Ultimately, this tool carries high translational potential to facilitate the physician's workflow and improve overall lagophthalmos patient care.
| Item Type: | Article |
|---|---|
| Uncontrolled Keywords: | TARSORRHAPHY; |
| Subjects: | 600 Technology > 610 Medical sciences Medicine |
| Divisions: | Medicine > Zentren des Universitätsklinikums Regensburg > Zentrum für Plastische-, Hand- und Wiederherstellungschirurgie |
| Depositing User: | Dr. Gernot Deinzer |
| Date Deposited: | 13 Mar 2024 09:44 |
| Last Modified: | 13 Mar 2024 09:44 |
| URI: | https://pred.uni-regensburg.de/id/eprint/59872 |
Actions (login required)
 |
View Item |
