Automated quantification of parafoveal microvascular complexity using deep learning-based OCTA segmentation and foveal avascular zone detection.

Multitask learning approach for foveal avascular zone segmentation in OCTA images

Identification and localization of myopic macular neovascularization using deep learning.

PURPOSE Myopia is rapidly increasing worldwide, with macular neovascularization (MNV) representing a major cause of vision loss. Early diagnosis remains challenging. Our study aims to incorporate a deep learning (DL) model to identify and localize myopic MNV using OCT images. METHODS A total of 6,215 OCT B-scan images of patients with pathologic myopia were included, of which 1,664 contained MNV. MNV lesions were manually annotated and validated by retina specialists. The dataset included additional myopia-related pathologies to account for factors affecting model robustness. Data were split into training and validation sets using patient-level 5-fold cross-validation. A YOLOv11 (You Only Look Once) object detection model was trained to localize MNV. Two postprocessing pipelines using either non-maximum suppression (NMS) or weighted box fusion (WBF) were tested. Model performance was evaluated at 25% intersection over union (IoU) using precision, recall, and mean average precision (mAP25). RESULTS The YOLOv11 models achieved mean precision ranging from 0.73 to 0.79 and recall from 0.76 to 0.80. The highest mAP25 reached 0.84. Adapting the model's output into an image-level binary classification yielded precision up to 0.85, recall up to 0.88, and F1-score up to 0.85. CONCLUSION A YOLOv11-based deep learning model reliably detects and localizes myopic MNV on OCT imaging, overcoming the challenge of structurally complex eyes with coexisting myopic pathologies. This tool may serve as an automated reader to assist with earlier diagnosis and timely anti-VEGF intervention in patients with pathologic myopia.

Annotated dataset for retinal fundus image segmentation: vascular arcade and optic nerve head

Digital fundoscopy is an ophthalmological technique used to evaluate the posterior segment of the eye, including the retina and its anatomical structures [1]. This technique acquires fundus images, visually documenting the aforementioned structures. This dataset was generated from a subset of images from the public Asia Pacific Tele-Ophthalmology Society 2019 Blindness Detection (APTOS 2019 BD) dataset [2], these were manually annotated to generate segmentation masks of anatomical fundus structures, specifically the vascular arcade and the optic nerve head. We contribute to the development and evaluation of artificial intelligence models under comparable conditions by generating data to serve as input for segmentation models, thereby supporting the design of more efficient algorithms for the automated identification of ophthalmological structures [3, 4]. This dataset consists of 500 masks generated from 500 images randomly selected from the APTOS 2019 BD dataset [2]. These are non-homogeneous color fundus images containing global labels for pathological alterations, illumination, as well as varying image resolutions. Image dimensions vary from 474 × 358 pixels to 4288 × 2848 pixels. This dataset provides manual masks for specific ophthalmic structures.

Deformable Models-Based Retinal OCT Layer Segmentation and Classification with Feature Analysis

Retinal Vessel Segmentation via Morphological Refinement and Adaptive Late Fusion

Retinal vessel segmentation is fundamental for quantitative retinal vascular analysis; however, accurate delineation remains challenging because of nonuniform illumination, low-contrast capillaries, pathological lesions, and variations in image acquisition. This study presents a rule-based unsupervised retinal vessel segmentation framework based on decision-level adaptive late fusion, in which six complementary local adaptive thresholding methods are integrated and subsequently refined through luminance validation, hysteresis reconstruction, component cleanup, elongation filtering, and boundary smoothing. In this context, unsupervised denotes that no statistical segmentation model is trained using manual vessel annotations; instead, fixed parameters and fusion weights are determined from a small development subset, while all evaluation images remain unseen during method development. The proposed framework was evaluated on 135 independent retinal fundus images from the DRIVE, STARE, CHASE_DB1, HRF, and LES-AV datasets using an eroded field-of-view protocol. It achieved an image-weighted Dice coefficient of 0.7104 (95% bootstrap confidence interval: 0.7005–0.7205), an IoU of 0.5539, a sensitivity of 0.7544, a specificity of 0.9597, a balanced accuracy of 0.8570, and a clDice score of 0.7395. Compared with fixed majority voting, the proposed adaptive late fusion strategy significantly improved segmentation performance in 128 of 135 test images, yielding a mean Dice improvement of 0.0270 (paired Wilcoxon, Holm-adjusted p = 7.67 × 10⁻²¹). Although segmentation performance remains limited for complex pathological images, particularly in the HRF dataset, the proposed framework demonstrates consistent cross-dataset performance, transparent decision-making, and strong reproducibility, providing an effective alternative when interpretable, training-free retinal vessel segmentation is required.

Deep Learning-based Automated Vessel Extraction for VR Image Creation From Cerebral TOF-MRA.

RATIONALE AND OBJECTIVES Accurate and efficient three-dimensional visualization of cerebral vasculature is essential for clinical evaluation; however, manual vessel extraction from time-of-flight (TOF) magnetic resonance angiography angiography (MRA) is time-consuming and operator-dependent. This study aimed to develop a deep learning-based cerebrovascular segmentation model and an automated vessel extraction method, and to evaluate their accuracy, volumetric reliability, and impact on volume rendering (VR) workflow efficiency. MATERIALS AND METHODS A 3D U-Net-based vessel segmentation model was trained using TOF-MRA images. Automated vessel extraction was performed by dilating predicted vessel regions by one voxel. Forty-eight intracranial aneurysm cases were analyzed. Segmentation performance was evaluated using the Dice similarity coefficient (DSC), normalized surface Dice (NSD); tolerance = 1 mm), and centerline distance (CLD). Inter-rater reliability was assessed using DSC between independently generated vessel masks in a subset of the dataset. Aneurysm volumes from original and vessel-extracted images were compared using equivalence testing with a 1% margin and two one-sided tests (TOST). VR image creation time was measured by 12 radiological technologists. RESULTS The DSC between independently generated vessel masks was 0.916. The DSC, recall, and precision of dilated vessel masks were significantly higher than those of non-dilated masks (p < 0.0001). The NSD was 0.982 ± 0.015, and the CLD was 0.196 ± 0.182 mm. Aneurysm volumes showed strong correlation (r = 0.999) with a small mean absolute error (MAE) (0.0915 mm³), and equivalence by TOST (p < 0.001). VR image creation time was significantly reduced (p = 0.0130). CONCLUSION The proposed method enables accurate, reproducible, and time-efficient generation of cerebral vascular VR images, suggesting its potential utility in clinical practice.