Jiong Zhang

Background: Retinal microvascular alterations observed in optical coherence tomography angiography (OCTA) have emerged as promising biomarkers for cerebral small vessel disease (CSVD). However, accurate quantification of OCTA-derived vascular biomarkers remains challenging due to imaging noise, projection artifacts, low contrast, and structural discontinuities, which often lead to unreliable vessel segmentation and distorted vascular topology. Purpose: To address these challenges, we propose SemTopNet, a semantic‐topology joint modeling framework for OCTA vessel segmentation. Methods: The proposed framework integrates vision‐language semantic guidance, centerline-guided topology refinement, and uncertainty-aware adaptive supervision within a unified architecture. Specifically, semantic embeddings are incorporated to enhance global structural representation, while a centerline-guided topology refinement module explicitly improves vascular continuity and suppresses fragmented predictions. In addition, Monte Carlo dropout‐based uncertainty estimation is employed to improve robustness in ambiguous and low signal‐to‐noise regions. Results: Extensive experiments on a CSVD OCTA dataset demonstrate that SemTopNet consistently outperforms existing convolutional and Transformer‐based segmentation methods, achieving the best performance in Dice similarity coefficient (0.8931), clDice (0.9027), Recall (0.9293), and AUC (0.9567). Ablation studies further verify the effectiveness of semantic modeling, topology‐aware refinement, and uncertainty‐aware supervision. Based on the obtained segmentation results, quantitative retinal vascular biomarkers were further analyzed, revealing significant layer-dependent microvascular alterations in CSVD patients, particularly in the deep vascular complex (DVC). Conclusion: These findings demonstrate that topology-consistent OCTA vessel segmentation is essential for reliable retinal microvascular quantification. The proposed SemTopNet provides a robust framework for OCTA biomarker extraction and retinal‐cerebral vascular association analysis in CSVD‐related studies.

Age-related macular degeneration (AMD) is a leading cause of vision impairment and blindness in the elderly, making accurate diagnosis and fine-grained sub type classification essential for clinical decision-making. However, existing methods often fail to fully leverage complementary information from multiple imaging modalities due to complex lesion characteristics and heterogeneous data distributions. To address these challenges, we pro pose a novel Multimodality Distillation and Fusion Network (ModiF) for robust AMD classification. First, modality specific feature extractors are designed to capture hierarchical semantic representations from different imaging modalities such as OCT, OCTA, and CFP. To further enhance feature discriminability, an auxiliary cross-modal soft supervision strategy is introduced during training, in which modality-specific predictions are adaptively voted and used as soft targets for knowledge distillation. Building on these modality-specific features, a principal-modality guided distillation mechanism is designed to facilitate effective cross-modality knowledge transfer while preserving critical diagnostic information. Furthermore, a confidence aware fusion module is introduced to adaptively balance the contributions of different modalities by dynamically adjusting fusion weights based on the entropy of modality specific predictions. Experimental validations on one internal dataset (OCT and OCTA) and two public datasets (CFP andOCT)demonstrate that ModiF consistently outperforms existing state-of-the-art methods across different settings, highlighting its effectiveness and generalization capability in AMD classification.

Background: The eye and brain share developmental and regulatory vascular similarities, and 3D vascular segmentation from Optical Coherence Tomography Angiography (OCTA) and Magnetic Resonance Angiography (MRA) is essential for neuro-ophthalmic assessment. However, vascular trees exhibit extreme multi-scale morphology and are vulnerable to topology disruptions caused by OCTA artifacts (e.g., noise and discontinuities) and MRA anatomical variants (e.g., Circle of Willis incompleteness), which often lead to fragmented vessels and unstable segment-wise predictions. Purpose: To develop a unified 3D vascular segmentation framework that transfers and shares structural priors across OCTA and MRA, enabling accurate parsing under extreme scale imbalance and topology disruptions, and producing anatomically plausible predictions that remain robust to imaging artifacts in OCTA and Circle-of-Willis variants in MRA. Methods: We propose GA-TAN, a Graph Aggregation and Topology-Aware Network for unified ocular–cerebral vascular segmentation. GA-TAN employs a variable-window hybrid CNN–Transformer backbone to adapt receptive fields across vessel scales, an Efficient Multi-Scale Attention (EMA) module to enhance volumetric spatial–channel interactions, and a Multi-Scale Graph Aggregation (MSGA) module to perform global topological reasoning and hierarchical feature fusion. A topology-aware training objective further supervises segment existence and connectivity to penalize structural discontinuities beyond pixel-wise losses. Experiments were conducted on Retina3D for binary 3D OCTA vessel segmentation and on TopCow for 13-class CoW artery segmentation from MRA. Results: GA-TAN achieved the best overall performance on Retina3D, with a DSC of 0.684 and a clDice of 0.667, indicating improved overlap accuracy and centerline connectivity preservation. On TopCow, GA-TAN obtained a mean Dice of 0.814 across 13 arterial categories and produced more anatomically consistent segment-wise predictions, showing clear advantages on thin communicating and variant-related segments. Ablation studies further validated the complementary contributions of EMA, MSGA, and topology-aware supervision. Conclusion: GA-TAN provides a unified and topology-aware solution for 3D vascular segmentation across OCTA and MRA. By integrating adaptive multi-scale modeling, graph-based global aggregation, and explicit topology supervision, the proposed framework improves both segmentation accuracy and vascular ontinuity, enhancing robustness to imaging artifacts and anatomical variants for ocular–cerebral vascular analysis.

Organoids, sophisticated in vitro models of human tissues, are crucial for medical research due to their ability to simulate organ functions and assess drug responses accurately. Accurate organoid instance segmentation is critical for quantifying their dynamic behaviors, yet remains profoundly limited by high-quality annotated datasets and pervasive overlap in microscopy imaging. While semi-supervised learning (SSL) offers a solution to alleviate reliance on scarce labeled data, conventional SSL frameworks suffer from biases induced by noisy pseudo-labels, particularly in overlapping regions. Synthesis-assisted SSL (SA-SSL) has been proposed for mitigating training biases in semi-supervised semantic segmentation. We present the first adaptation of SA-SSL to organoid instance segmentation and reveal that SA-SSL struggles to disentangle intertwined organoids, often misrepresenting overlapping instances as a single entity. To overcome this, we propose Pseudo-Label Unmixing (PLU), which identifies erroneous pseudo-labels for overlapping instances and then regenerates organoid labels through instance decomposition. For image synthesis, we apply a contour-based approach to synthesize organoid instances efficiently, particularly for overlapping cases. Instance-level augmentations (IA) on pseudo-labels before image synthesis further enhances the effect of synthetic data (SD). Rigorous experiments on two organoid datasets demonstrate our method's effectiveness, achieving performance comparable to fully supervised models using only 10% labeled data, and state-of-the-art results. Ablation studies validate the contributions of PLU, contour-based synthesis, and augmentation-aware training. By addressing overlap at both pseudo-label and synthesis levels, our work advances scalable, label-efficient organoid analysis, unlocking new potential for high-throughput applications in precision medicine.