Scholarlink

<p>Retinal image matching plays a pivotal role in medical diagnostics, longitudinal analysis of disease progression, and treatment planning. Accurate alignment of retinal images is especially crucial in multimodal scenarios where local structural differences and nonlinear radiometric differences pose significant challenges. While deep learning–based approaches have demonstrated promising results, they often suffer from data dependency issues. On the other hand, traditional handcrafted methods, predominantly based on SIFT-like descriptors, struggle to handle integrated geometric and radiometric differences and structural displacements, leaving significant room for improvement. To address these limitations, this paper proposes a straightforward yet robust descriptor that adapts to multiple multimodal and monomodal datasets, including color fundus (CF) and fluorescein angiography (FA), ultra-widefield FA, scanning laser ophthalmoscopy (SLO images), and the CF FIRE dataset with varying overlaps. The core of the proposed framework lies in an advanced feature space constructed from first- and second-order directional anisotropic texton-like Leung–Malik (LM) derivatives. Their elongated receptive fields align naturally with delicate vascular and anatomical structures, enabling more consistent feature encoding. The extracted feature maps are fused into a discriminative representation by leveraging the orientation index of the maximum directional response within each local patch. Eventually, a weighted adaptive binning configuration is designed to handle local variations. The proposed approach is scale-invariant, which is crucial for most datasets. To further enhance flexibility, we have incorporated a rotation-invariant module. Without any failures, the proposed method achieved an average recall of 37.48%. It outperformed state-of-the-art approaches, including steerable Gaussian derivatives, scale-invariant RIFT, HOSS, PIIFD, HAPCG, and recent deep methods such as RIPE and the HardNet family. Code is available from <a href="https://github.com/AmSedaghat/HLMF-descriptor">https://github.com/AmSedaghat/HLMF-descriptor</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust Retinal Image Matching: A Modality-Resistant Descriptor Using Directional Anisotropic Texton-Like Features and Evaluation Across Diverse Datasets

  • 期刊论文

  • 发表日期：

    2026年1月9日

查看全文
  • Negar Jovhari,
  • Amin Sedaghat,
  • Reza Shah-Hosseini,
  • Nazila Mohammadi

摘要

Retinal image matching plays a pivotal role in medical diagnostics, longitudinal analysis of disease progression, and treatment planning. Accurate alignment of retinal images is especially crucial in multimodal scenarios where local structural differences and nonlinear radiometric differences pose significant challenges. While deep learning–based approaches have demonstrated promising results, they often suffer from data dependency issues. On the other hand, traditional handcrafted methods, predominantly based on SIFT-like descriptors, struggle to handle integrated geometric and radiometric differences and structural displacements, leaving significant room for improvement. To address these limitations, this paper proposes a straightforward yet robust descriptor that adapts to multiple multimodal and monomodal datasets, including color fundus (CF) and fluorescein angiography (FA), ultra-widefield FA, scanning laser ophthalmoscopy (SLO images), and the CF FIRE dataset with varying overlaps. The core of the proposed framework lies in an advanced feature space constructed from first- and second-order directional anisotropic texton-like Leung–Malik (LM) derivatives. Their elongated receptive fields align naturally with delicate vascular and anatomical structures, enabling more consistent feature encoding. The extracted feature maps are fused into a discriminative representation by leveraging the orientation index of the maximum directional response within each local patch. Eventually, a weighted adaptive binning configuration is designed to handle local variations. The proposed approach is scale-invariant, which is crucial for most datasets. To further enhance flexibility, we have incorporated a rotation-invariant module. Without any failures, the proposed method achieved an average recall of 37.48%. It outperformed state-of-the-art approaches, including steerable Gaussian derivatives, scale-invariant RIFT, HOSS, PIIFD, HAPCG, and recent deep methods such as RIPE and the HardNet family. Code is available from https://github.com/AmSedaghat/HLMF-descriptor.