Iglesias Morís, DanielMoura, Joaquim deCarmona Suárez, Enrique JavierNovo, JorgeOrtega, Marcos2025-04-292025-04-292025-01-13Daniel I. Morís, Joaquim de Moura, Enrique J. Carmona, Jorge Novo, Marcos Ortega. Semantic-guided generative latent diffusion augmentation approaches for improving the neovascularization diagnosis in OCT-A imaging. Pattern Recognition Letters 189, 31-37, 2025. DOI: https://doi.org/10.1016/j.patrec.2025.01.0030167-8655https://doi.org/10.1016/j.patrec.2025.01.003https://hdl.handle.net/20.500.14468/26507The registered version of this article, first published in “Pattern Recognition Letters 189, 31-37, 2025", is available online at the publisher's website: Elsevier, https://doi.org/10.1016/j.patrec.2025.01.003 La versión registrada de este artículo, publicado por primera vez en “Pattern Recognition Letters 189, 31-37, 2025.", está disponible en línea en el sitio web del editor: Elsevier, https://doi.org/10.1016/j.patrec.2025.01.003Age-related Macular Degeneration (AMD) presents an enormous challenge in Western Societies due to the increase in life expectancy. AMD is characterized for causing Macular Neovascularization. Optical Coherence Tomography Angiography (OCT-A) represents an advanced method to help find evidence of the disease. In this context, deep learning algorithms are suitable to make a screening of the disease. However, biomedical imaging domains are usually affected by the data scarcity issue. The mitigation of this problem can be achieved with the support of generative latent diffusion models. This represents a powerful strategy to artificially augment the cardinality of the original dataset. In this work, we present a novel fully automatic methodology to generate OCT-A images, guided by semantic information, to reduce the impact of data scarcity and to enable an accurate neovascularization diagnosis. The evaluation has been performed with a specific dataset composed of two different fields of view commonly used by clinicians. The results demonstrate a top accuracy of 96.50% 1.37%, using 3 × 3 scans, and 95.79% 1.44%, when using 6 × 6 scans. The proposed methodology has great potential to be extrapolated to other imaging modalities and domains.eninfo:eu-repo/semantics/openAccess33 Ciencias Tecnológicasartículodeep learninglatent diffusionimage generationOCT-AAMDneovascularization