Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SwinVFTR: A Novel Volumetric Feature-learning Transformer for 3D OCT Fluid Segmentation

Khondker Fariha Hossain, Sharif Amit Kamran, Alireza Tavakkoli, George Bebis, Sal Baker

TL;DR

SwinVFTR targets precise fluid segmentation in 3D OCT volumes across multiple vendor devices by combining channel-wise volumetric sampling with a Swin-transformer encoder that uses shifted windows and a multi-receptive-field residual design. A volumetric attention skip mechanism and a 3D CNN-style decoder enable robust fusion of encoder-decoder features, improving boundary delineation for IRF, SRF, and PED across Cirrus, Spectralis, and Topcon data. Quantitative results show superior mean Dice, IOU, and SSIM versus state-of-the-art CNN and Transformer baselines, indicating strong generalization to cross-vendor OCT data. The approach offers a practical path toward accurate, scalable 3D fluid localization in ophthalmic imaging and can extend to other modalities.

Abstract

Accurately segmenting fluid in 3D optical coherence tomography (OCT) images is critical for detecting eye diseases but remains challenging. Traditional autoencoder-based methods struggle with resolution loss and information recovery. While transformer-based models improve segmentation, they arent optimized for 3D OCT volumes, which vary by vendor and extraction technique. To address this, we propose SwinVFTR, a transformer architecture for precise fluid segmentation in 3D OCT images. SwinVFTR employs channel-wise volumetric sampling and a shifted window transformer block to improve fluid localization. Moreover, a novel volumetric attention block enhances spatial and depth-wise attention. Trained using multi-class dice loss, SwinVFTR outperforms existing models on Spectralis, Cirrus, and Topcon OCT datasets, achieving mean dice scores of 0.72, 0.59, and 0.68, respectively, along with superior performance in mean intersection-over-union (IOU) and structural similarity (SSIM) metrics.

SwinVFTR: A Novel Volumetric Feature-learning Transformer for 3D OCT Fluid Segmentation

TL;DR

SwinVFTR targets precise fluid segmentation in 3D OCT volumes across multiple vendor devices by combining channel-wise volumetric sampling with a Swin-transformer encoder that uses shifted windows and a multi-receptive-field residual design. A volumetric attention skip mechanism and a 3D CNN-style decoder enable robust fusion of encoder-decoder features, improving boundary delineation for IRF, SRF, and PED across Cirrus, Spectralis, and Topcon data. Quantitative results show superior mean Dice, IOU, and SSIM versus state-of-the-art CNN and Transformer baselines, indicating strong generalization to cross-vendor OCT data. The approach offers a practical path toward accurate, scalable 3D fluid localization in ophthalmic imaging and can extend to other modalities.

Abstract

Accurately segmenting fluid in 3D optical coherence tomography (OCT) images is critical for detecting eye diseases but remains challenging. Traditional autoencoder-based methods struggle with resolution loss and information recovery. While transformer-based models improve segmentation, they arent optimized for 3D OCT volumes, which vary by vendor and extraction technique. To address this, we propose SwinVFTR, a transformer architecture for precise fluid segmentation in 3D OCT images. SwinVFTR employs channel-wise volumetric sampling and a shifted window transformer block to improve fluid localization. Moreover, a novel volumetric attention block enhances spatial and depth-wise attention. Trained using multi-class dice loss, SwinVFTR outperforms existing models on Spectralis, Cirrus, and Topcon OCT datasets, achieving mean dice scores of 0.72, 0.59, and 0.68, respectively, along with superior performance in mean intersection-over-union (IOU) and structural similarity (SSIM) metrics.
Paper Structure (14 sections, 2 equations, 2 figures, 1 table)

This paper contains 14 sections, 2 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Channel-wise Volumetric Sampling
  4. Proposed Swin-Transformer Block
  5. Encoder
  6. Volumetric Attention Block
  7. Decoder
  8. Experiments
  9. Dataset and Preprocessing
  10. Hyper-parameter Initialization
  11. Quantitative Evaluation
  12. Conclusion
  13. ACKNOWLEDGEMENTS
  14. COMPLIANCE WITH ETHICAL STANDARDS

Figures (2)

  • Figure 1: Proposed SwinVFTR architecture which takes 3D OCT volume input with channel-wise sampling technique and outputs a 3D segmentation map of the fluid accumulation. The SwinVFTR encoder incorporates a new swin-transformer block consisting of Shifted window attention, Multi-headed attention and Multi-Receptive Field (MRF) sub-block with both convolution and dilated convolution layers. The encoder features are sequentially added with the decoder using a skip-connection consisting of a volumetric attention (VA) block.
  • Figure 2: SwinVFTR segments fluid with better precision than other 3D CNN and Transformer architectures. The row contains Cirrus, Spectralis and Topcon data-sets. Whereas the column contains ground-truths and segmentation maps for SwinVFTR, SwinUNETR, UNETR, ResUNet-3D and Attention-UNet-3D. Here, IRF, SRF, and PED fluids are colored as Red, Yellow and Blue.