Optimizing deep learning models for glaucoma screening with vision transformers for resource efficiency and the pie augmentation method

Sirikorn Sangchocanonta, Pakinee Pooprasert, Nichapa Lerthirunvibul, Kanyarak Patchimnan, Phongphan Phienphanich, Adirek Munthuli, Sujittra Puangarom, Rath Itthipanichpong, Kitiya Ratanawongphaibul, Sunee Chansangpetch, Anita Manassakorn, Visanee Tantisevi, Prin Rojanapongpun and Charturong Tantibundhit

PLOS ONE, 2025, vol. 20, issue 3, 1-28

Abstract: Glaucoma is the leading cause of irreversible vision impairment, emphasizing the critical need for early detection. Typically, AI-based glaucoma screening relies on fundus imaging. To tackle the resource and time challenges in glaucoma screening with convolutional neural network (CNN), we chose the Data-efficient image Transformers (DeiT), a vision transformer, known for its reduced computational demands, with preprocessing time decreased by a factor of 10. Our approach utilized the meticulously annotated GlauCUTU-DATA dataset, curated by ophthalmologists through consensus, encompassing both unanimous agreement (3/3) and majority agreement (2/3) data. However, DeiT’s performance was initially lower than CNN. Therefore, we introduced the “pie method," an augmentation method aligned with the ISNT rule. Along with employing polar transformation to improved cup region visibility and alignment with the vision transformer’s input to elevated performance levels. The classification results demonstrated improvements comparable to CNN. Using the 3/3 data, excluding the superior and nasal regions, especially in glaucoma suspects, sensitivity increased by 40.18% from 47.06% to 88.24%. The average area under the curve (AUC) ± standard deviation (SD) for glaucoma, glaucoma suspects, and no glaucoma were 92.63 ± 4.39%, 92.35 ± 4.39%, and 92.32 ± 1.45%, respectively. With the 2/3 data, excluding the superior and temporal regions, sensitivity for diagnosing glaucoma increased by 11.36% from 47.73% to 59.09%. The average AUC ± SD for glaucoma, glaucoma suspects, and no glaucoma were 68.22 ± 4.45%, 68.23 ± 4.39%, and 73.09 ± 3.05%, respectively. For both datasets, the AUC values for glaucoma, glaucoma suspects, and no glaucoma were 84.53%, 84.54%, and 91.05%, respectively, which approach the performance of a CNN model that achieved 84.70%, 84.69%, and 93.19%, respectively. Moreover, the incorporation of attention maps from DeiT facilitated the precise localization of clinically significant areas, such as the disc rim and notching, thereby enhancing the overall effectiveness of glaucoma screening.

Date: 2025
References: View complete reference list from CitEc
Citations:

Downloads: (external link)
https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0314111 (text/html)
https://journals.plos.org/plosone/article/file?id= ... 14111&type=printable (application/pdf)

Related works:
This item may be available elsewhere in EconPapers: Search for items with the same title.

Export reference: BibTeX RIS (EndNote, ProCite, RefMan) HTML/Text

Persistent link: https://EconPapers.repec.org/RePEc:plo:pone00:0314111

DOI: 10.1371/journal.pone.0314111

Access Statistics for this article

More articles in PLOS ONE from Public Library of Science
Bibliographic data for series maintained by plosone ().