Dense-UNet++ with CBAM and ASPP for Automated Vitreous Hyperreflective Foci Segmentation in OCT

• Published in: 2025 International Conference on Sensors and Related Networks (SENNET) Special Focus on Digital Healthcare(64220) pp. 1 - 6
• Main Authors: S V, Adithiya, G, Dharani Bai
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.07.2025
• Subjects: ASPP; Attention Mechanism; Attention mechanisms; Biomedical imaging; CBAM; Dense Blocks; Image segmentation; Manuals; Medical Image Segmentation; Optical coherence tomography; Optical Coherence Tomography (OCT); Optical sensors; Retina; Spectral analysis; Three-dimensional displays; Tuning; UNet; Vitreous Hyperreflective Foci (VHF)
• DOI: 10.1109/SENNET64220.2025.11136056

Vitreous Hyperreflective Foci (vHRF) detected by Spectral Domain Optical Coherence Tomography (SD-OCT) are an essential imaging biomarker for diagnosing and monitoring inflammation in the posterior region, especially in patients with uveitis macular edema. Manual identification of vHRF is time-consuming and subject to variability between observers. This paper presents a deep learning framework for the automated segmentation of vHRF, utilising a Dense-UNet++ architecture that incorporates dense blocks, Atrous Spatial Pyramid Pooling (ASPP), and Convolutional Block Attention Modules (CBAM). This network improves the reuse of features, captures the multiscale context, and emphasises relevant spatial and channel-wise features through attention mechanisms. This proposed model is trained on a hospital dataset of OCT B-scans with expert-level manual annotated vHRF masks, the model achieved a Dice Similarity Coefficient of 0.81 ± 0.12 on the test set. The findings highlight the ability of the architecture to precisely segment very small, sparse vHRF regions, suggesting that it may find use in clinical settings for the objective assessment of inflammation.