Book chapter
abVAE: Attribute-Based Booster Variational Autoencoder for Interpretable Latent Presentation in Optical Coherence Tomography of Glaucomatous Eyes
Ophthalmic Medical Image Analysis, pp.137-146
Lecture Notes in Computer Science, v. 16209, Springer Nature Switzerland
2025
DOI: 10.1007/978-3-032-10351-2_14
Abstract
Glaucoma is a chronic optic neuropathy characterized by progressive retinal ganglion cell loss. To better visualize glaucomatous spatial patterns of nerve loss, we propose the attribute-based booster variational autoencoder (abVAE), which enables controllable latent representations without compromising reconstruction performance. Building upon the booster VAE (bVAE) framework and inspired by the attribute alignment loss introduced in Attri-VAE [1], the abVAE preserves reconstruction fidelity while enabling attribute-specific controllability over the inferior temporal (IT) and superior temporal (ST) sectors within the elliptical annulus of the retinal ganglion cell plus inner plexiform layer (GCIPL) thickness map from optical coherence tomography scans. By design, the latent space montage maps reveal that thicker regions are concentrated in the upper right corner, while thinner regions appear in the lower left. Quantitatively, the linear relationships between the latent variables and anatomical attributes (d1 $$d_1$$ –TIT∗ $$T^*_{IT}$$ and d2 $$d_2$$ –TST∗ $$T^*_{ST}$$ ) are reflected in the mean values of R2 $$R^2$$ : 0.95±0.01 $$0.95\pm 0.01$$ and 0.86±0.04 $$0.86\pm 0.04$$ for the abVAE model, 0.76±0.06 $$0.76\pm 0.06$$ and 0.45±0.22 $$0.45\pm 0.22$$ for the bVAE model, and 0.64±0.21 $$0.64\pm 0.21$$ and 0.23±0.32 $$0.23\pm 0.32$$ for the β $$\beta $$ -VAE model. The model also achieves high reconstruction quality, with a Dice score of 0.99 and a structural similarity index (SSIM) of 0.73. These results demonstrate that abVAE effectively balances anatomical interpretability and reconstruction accuracy, making it suitable for modeling spatial patterns of retinal thinning in glaucoma.
Details
- Title: Subtitle
- abVAE: Attribute-Based Booster Variational Autoencoder for Interpretable Latent Presentation in Optical Coherence Tomography of Glaucomatous Eyes
- Creators
- Pei-Hsin Chiu - University of IowaBrett A. Johnson - University of IowaEdward F. Linton - University of IowaAndrew E. Pouw - University of IowaMichael Wall - University of IowaYoung H. Kwon - University of IowaRandy H. Kardon - University of IowaJui-Kai Wang - The University of Texas Southwestern Medical CenterMona K. Garvin - Iowa City VA Health Care System
- Contributors
- Huihui Fang (Editor)Meng Wang (Editor)Heng Li (Editor)Hao Chen (Editor)Hrvoje Bogunović (Editor)Cecilia S. Lee (Editor)
- Resource Type
- Book chapter
- Publication Details
- Ophthalmic Medical Image Analysis, pp.137-146
- Series
- Lecture Notes in Computer Science; v. 16209
- DOI
- 10.1007/978-3-032-10351-2_14
- eISSN
- 1611-3349
- ISSN
- 0302-9743
- Publisher
- Springer Nature Switzerland; Cham
- Language
- English
- Date published
- 2025
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering; Neurology; Electrical and Computer Engineering; Iowa Neuroscience Institute; Ophthalmology and Visual Sciences
- Record Identifier
- 9985033950702771
Metrics
1 Record Views