Scholarlink

Background <p>Diabetic retinopathy (DR) is a leading cause of visual impairment worldwide. Early and accurate classification of disease severity using retinal fundus images is critical for timely diagnosis and prevention of vision loss. Automated systems can assist clinicians by improving screening efficiency and reducing diagnostic variability.</p> Objective <p>This study proposes a novel feature-pitched classification model (FPCM) for automated classification of diabetic retinopathy stages using color fundus images, aiming to improve accuracy through enhanced feature extraction and spatial attention mechanisms.</p> Methods <p>The proposed FPCM model extracts texture, intensity, and structural features from fundus images by identifying high-information regions referred to as pitch points. These features are processed using a concentric transformer learning (CTL) mechanism, which applies spatial attention across nested regions to capture both local and global patterns. The model was evaluated using the Kaggle Diabetic Retinopathy dataset. Performance metrics included accuracy, precision, and sensitivity. Statistical reliability was assessed using bootstrap-based confidence interval analysis, and results were compared with existing methods such as ERCN, EffNet-SVM, HPLBO_DMN, FCSAM, CLAHE TH, and MSTNet.</p> Results <p>The proposed model achieved an accuracy of 95.19%, precision of 95.66%, and sensitivity of 95.29%. Bootstrap analysis confirmed the robustness and statistical reliability of the performance. Comparative evaluation demonstrated that FPCM provides competitive or superior results relative to existing approaches.</p> Conclusion <p>The proposed FPCM model effectively enhances diabetic retinopathy severity classification by integrating pitch-point–based feature extraction with concentric transformer learning. The model demonstrates high accuracy and robustness while focusing on clinically relevant regions, improving both interpretability and diagnostic reliability. These findings highlight the potential of FPCM for deployment in automated retinal screening systems, although further validation on diverse clinical datasets is recommended.</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

Feature pitched transformer model for automated diabetic retinopathy severity classification on fundus images

  • 期刊论文

  • 发表日期：

    2026年5月2日

查看全文
  • Allin Geo A.V.,
  • Lavanya M.,
  • Govindharaj I.,
  • Malathi V.

摘要

Background

Diabetic retinopathy (DR) is a leading cause of visual impairment worldwide. Early and accurate classification of disease severity using retinal fundus images is critical for timely diagnosis and prevention of vision loss. Automated systems can assist clinicians by improving screening efficiency and reducing diagnostic variability.

Objective

This study proposes a novel feature-pitched classification model (FPCM) for automated classification of diabetic retinopathy stages using color fundus images, aiming to improve accuracy through enhanced feature extraction and spatial attention mechanisms.

Methods

The proposed FPCM model extracts texture, intensity, and structural features from fundus images by identifying high-information regions referred to as pitch points. These features are processed using a concentric transformer learning (CTL) mechanism, which applies spatial attention across nested regions to capture both local and global patterns. The model was evaluated using the Kaggle Diabetic Retinopathy dataset. Performance metrics included accuracy, precision, and sensitivity. Statistical reliability was assessed using bootstrap-based confidence interval analysis, and results were compared with existing methods such as ERCN, EffNet-SVM, HPLBO_DMN, FCSAM, CLAHE TH, and MSTNet.

Results

The proposed model achieved an accuracy of 95.19%, precision of 95.66%, and sensitivity of 95.29%. Bootstrap analysis confirmed the robustness and statistical reliability of the performance. Comparative evaluation demonstrated that FPCM provides competitive or superior results relative to existing approaches.

Conclusion

The proposed FPCM model effectively enhances diabetic retinopathy severity classification by integrating pitch-point–based feature extraction with concentric transformer learning. The model demonstrates high accuracy and robustness while focusing on clinically relevant regions, improving both interpretability and diagnostic reliability. These findings highlight the potential of FPCM for deployment in automated retinal screening systems, although further validation on diverse clinical datasets is recommended.