Scholarlink

Cardiac segmentation is essential for accurate diagnosis and treatment planning in cardiovascular diseases. Deep learning methods, particularly convolutional neural networks (CNNs), have improved the efficiency of segmentation by capturing local spatial features. However, transformers offer better modeling of long-range dependencies and global context, making them suitable for image processing tasks. This study utilizes the UNet Transformer (UNETR) architecture for cardiac segmentation on the Automatic Cardiac Diagnosis Challenge (ACDC) dataset, which includes cine-MRI images of the heart. The dataset consisted of 150 patients, with manual annotations for three semantic classes: left ventricle (LV), myocardium (MYO), and right ventricle (RV). The model was fine-tuned on the ACDC dataset with pre-trained ImageNet weights. UNETR’s integration of transformers into the traditional UNet framework shows potential; however, data-hungry transformer-based structure does not capture complex dependencies with a smaller size dataset. Through extensive experiments and comparison with foundational models (Segment Anything Model (SAM) and MedSAM), our approach demonstrates competitive performance and generalization capabilities.

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

Cardiac Segmentation Using UNETR: A Transformer Based Deep Learning Approach on the ACDC Dataset

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Tijana Geroski,
  • Ognjen Pavić,
  • Lazar Dašić,
  • Amir Amini,
  • Nenad Filipović

摘要

Cardiac segmentation is essential for accurate diagnosis and treatment planning in cardiovascular diseases. Deep learning methods, particularly convolutional neural networks (CNNs), have improved the efficiency of segmentation by capturing local spatial features. However, transformers offer better modeling of long-range dependencies and global context, making them suitable for image processing tasks. This study utilizes the UNet Transformer (UNETR) architecture for cardiac segmentation on the Automatic Cardiac Diagnosis Challenge (ACDC) dataset, which includes cine-MRI images of the heart. The dataset consisted of 150 patients, with manual annotations for three semantic classes: left ventricle (LV), myocardium (MYO), and right ventricle (RV). The model was fine-tuned on the ACDC dataset with pre-trained ImageNet weights. UNETR’s integration of transformers into the traditional UNet framework shows potential; however, data-hungry transformer-based structure does not capture complex dependencies with a smaller size dataset. Through extensive experiments and comparison with foundational models (Segment Anything Model (SAM) and MedSAM), our approach demonstrates competitive performance and generalization capabilities.