This paper presents AFMUNet, an Adaptive Filter-Based Frequency Modulation U-Net for Optical Coherence Tomography Angiography (OCTA) segmentation. The model addresses the challenge of segmenting both small blood vessels and larger vascular structures, which exhibit significant variations in scale, contrast, and connectivity. To tackle these issues, AFMUNet achieves an image-sized receptive field while capturing global dependencies, enhancing performance across diverse vessel sizes. Specifically, it incorporates the Fast Fourier Transform (FFT) applied to feature maps across multiple scales of a UNet-like architecture. This component is crucial for identifying critical global frequency patterns, enabling the model to represent the intricate details of small vessels alongside larger ones. Simultaneously, a lightweight attention block is employed to learn adaptive frequency filters from the FFT-derived representations. This mechanism selectively emphasizes transferable frequency components essential for highlighting small and thin vessels while suppressing noise and less informative features that could detract from the segmentation of large vascular structures. Experimental evaluations demonstrate that AFMUNet outperforms state-of-the-art models, achieving mean Dice and mean Intersection over Union (IoU) scores of 91.69% and 84.65%, respectively. These results underscore its robustness and superior ability to accurately segment OCTA images, effectively addressing the dual challenge of capturing both fine-grained and coarse vascular details.

This paper introduces a novel framework, termed RAW (Region Attention-Weighted Network with Inter-Region Information Exchange), specifically developed for Age-related Macular Degeneration (AMD) grading using high-resolution input images. The framework begins with a High-Fidelity Detail Retention (HFDR) block, designed to preserve critical image details essential for accurate analysis. Furthermore, a lightweight network, inspired by ResNet, is incorporated to facilitate efficient feature extraction. This is augmented by a Region Attention-Weighted (RAW) block, which highlights significant regions while mitigating interference from less relevant areas. To ensure the effective propagation of crucial information for precise AMD grading, we developed the Inter-Region Information Exchange (IRIE) block. The efficacy of the proposed RAW framework was evaluated using a dataset comprising 864 Color Retinal Fundus (CRF) images, demonstrating outstanding performance compared to existing methods. It achieved a mean accuracy of 98.08%, a mean F1-score of 98.06%, and a mean Cohen’s Kappa of 97.42%. Statistical analysis further confirms the framework’s significant advantage, particularly in terms of F1-score, underscoring its robustness and exceptional capability for AMD grading.

This paper introduces CrossDR, a novel approach unifying 2D and 3D feature representations through a context-aware cross-attention mechanism for diabetic retinopathy (DR) classification in 3D Optical Coherence Tomography (OCT) scans. The model uncovers critical diagnostic cues embedded in 3D-OCT images, which offer rich information for DR detection. The framework comprises three core components: the Lightweight Attention (LA) block, a CNN-based encoder, and the Context-Aware Cross-Attention (CA)2block. The LA block highlights volume-specific features that are crucial for DR diagnosis. It achieves this by applying a series of 1×1 convolutions followed by softmax operations, enabling the extraction of spatially informative DR-specific features. Meanwhile, the CNN-based encoder extracts high-level semantic features from the 3D-OCT slices, offering a robust representation of retinal structures. The (CA)2block further boosts the model’s performance by dynamically capturing and reweighting feature dependencies. This block models relationships between feature maps at different abstraction levels, improving the discriminative power of the extracted features. By integrating these refined features through the (CA)2block, the model achieves accurate DR classification. The CrossDR framework was evaluated on 481 volumetric OCT images, outperforming existing state-of-the-art methods with an accuracy of 94%. Multiple experiments and ablation studies were conducted to assess our model’s performance.