Comparative Analysis of Retinal Fluid Segmentation in OCT Images : Evaluating CNN-Based Models vs. Transformer-Based Models

Semantic segmentation of retinal fluids in optical coherence tomography (OCT) images plays a crucial role in the diagnosis and treatment of retinal diseases. This study presents a comprehensive comparative analysis of transformer-based models and convolutional neural network (CNN)-based models for retinal fluid segmentation, focusing the comparison between the CNN-based models, DeepLab and UNet, and the transformer-based model, Swin UNetR. The evaluation metrics employed include F1 score and Absolute Volume Difference (AVD). The RETOUCH challenge dataset from three different OCT vendors, Spectralis, Cirrus, and Topcon, was used for evaluating the models. The results indicate that Swin UNetR outperformed the other models, achieving the best segmentation results for both F1 score and AVD. Visual analysis further supported these findings, revealing Swin UNetR's superior capability in capturing intricate fluid structures. Statistical analysis using paired t-tests and ANOVA showed significance between Swin UNetR and UNet for the Serous Retinal Fluid (SRF) class. The combination of the Swin Transformer and U-Net architecture in Swin UNetR proved to be a potent approach, capturing both global contextual information and precise local details. This research provides valuable insights into the strengths and weaknesses of different segmentation methods, guiding future advancements in the field of retinal fluid segmentation and contributing to improved clinical diagnosis and treatment of retinal diseases.