Abstract
Purpose :
Optical Coherence Tomography (OCT) has emerged as a leading tool for the automatic diagnosis of ocular diseases. However, depending on the targeted structures (such as the macula, optic nerve head, etc.) and the device used, the number of scans and the inter-slice spacing may vary. This variation leads to variable anisotropic volumes, posing challenges for automated diagnosis. This study aims to tackle these challenges and enhance multiple disease screening using OCT.
Methods :
We collected 663 OCT volumes from 251 patients seeking vision-related consultation using the Heidelberg Spectralis device at Brest University Hospital in France between 2017 and 2018. An ophthalmologist interpreted these OCT and identified 7 distinct diseases. We proposed a novel transformer-based method named IsotropicViT. This method operates on patches of constant size in cubic millimeters, as opposed to patches of constant size in voxels, as generally done in the literature. The aim is to achieve isotropic 3D representation. We conducted a comparison between this method and the standard 3D Vision Transformer (ViT), which operates on patches of constant size in voxels, on normal (205 volumes) versus pathological (458 volumes) classification. We perform a 4-fold cross-validation on 80% of the data. The resulting ensemble of 4 models was evaluated by the area under the receiver operating characteristic curve (AUC) on the test set (20%). The Delong test was used for statistical analysis.
Results :
The OCT volumes displayed varying scan numbers, ranging from 19 to 62 (mean=24, sd=10), and inter-slice spacing between 0.03 to 0.29 millimeters (mean=0.23, sd=0.05). Using the ViT ensemble, we achieved an AUC = 0.878, while the IsotropicViT ensemble reached an AUC = 0.936. Notably, our approach showcased a statistically significant improvement (p-value = 0.02 < 0.05).
Conclusions :
Our proposed method demonstrated considerable improvements by emphasizing specific patch sizes and flexible data processing. These findings highlight the importance of addressing issues related to anisotropy and varying data sizes. Our next step involves exploring self-supervised learning to pretrain the classification models for improved performance.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.