Scholarlink

<p>Cardiac magnetic resonance imaging (MRI) plays a key role in the diagnosis of cardiovascular diseases, yet manual segmentation remains time-consuming and subject to observer variability. This study presents a fully automated deep learning framework for accurate segmentation of the left ventricular endocardium and epicardium, enabling the extraction of clinically relevant geometric indices. The proposed method is based on a Fully Convolutional Network (FCN) with skip connections and incorporates a hybrid loss function combining contour alignment, level-set regularization, and gradient-aware constraints to improve boundary precision while maintaining a compact architecture (10.9&#xa0;M parameters).The model was trained and evaluated on the ACDC dataset, and a dedicated quantification pipeline was developed to compute left ventricular cavity area, myocardial area, dimensions, and regional wall thickness. As the ACDC dataset does not provide ground-truth clinical indices, reference measurements were generated using a semi-automated approach cross-validated with ImageJ, and further evaluation was conducted on the LVQuan2018 dataset.The proposed framework achieved competitive segmentation performance, with Dice scores of 0.91 ± 0.04 for the endocardium and 0.93 ± 0.03 for the epicardium. Quantification results demonstrated good agreement with reference measurements across all indices, particularly for cavity and myocardial areas and regional wall thickness, while expected discrepancies were observed for dimensional estimation due to methodological differences.By integrating boundary-aware segmentation with interpretable quantification, the proposed framework provides an efficient end-to-end tool for cardiac assessment and supporting the evaluation of structural abnormalities such as hypertrophic cardiomyopathy and heart failure.</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

Automated left ventricle segmentation using FCNs for quantitative assessment of cardiac indices in cardiovascular disease

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Samira Jaballah,
  • Imen Baklouti,
  • Tarak Ben Said,
  • Mohamed Ben Slima

摘要

Cardiac magnetic resonance imaging (MRI) plays a key role in the diagnosis of cardiovascular diseases, yet manual segmentation remains time-consuming and subject to observer variability. This study presents a fully automated deep learning framework for accurate segmentation of the left ventricular endocardium and epicardium, enabling the extraction of clinically relevant geometric indices. The proposed method is based on a Fully Convolutional Network (FCN) with skip connections and incorporates a hybrid loss function combining contour alignment, level-set regularization, and gradient-aware constraints to improve boundary precision while maintaining a compact architecture (10.9 M parameters).The model was trained and evaluated on the ACDC dataset, and a dedicated quantification pipeline was developed to compute left ventricular cavity area, myocardial area, dimensions, and regional wall thickness. As the ACDC dataset does not provide ground-truth clinical indices, reference measurements were generated using a semi-automated approach cross-validated with ImageJ, and further evaluation was conducted on the LVQuan2018 dataset.The proposed framework achieved competitive segmentation performance, with Dice scores of 0.91 ± 0.04 for the endocardium and 0.93 ± 0.03 for the epicardium. Quantification results demonstrated good agreement with reference measurements across all indices, particularly for cavity and myocardial areas and regional wall thickness, while expected discrepancies were observed for dimensional estimation due to methodological differences.By integrating boundary-aware segmentation with interpretable quantification, the proposed framework provides an efficient end-to-end tool for cardiac assessment and supporting the evaluation of structural abnormalities such as hypertrophic cardiomyopathy and heart failure.