Scholarlink

<p>The function of the heart left ventricle (LV) plays a critical role within the human circulatory system and can be analyzed with SPECT images. Due to the inherent nature of SPECT images being low-resolution, noisy, and low-contrast, myocardial segmentation becomes a challenging task. Manual processes such as tightly cropping the heart region become necessary to partially eliminate part of these issues. For cardiac images without problematic regions, non–AI-based approaches utilizing shape priors on manually cropped images perform sufficiently well; however, their performance significantly deteriorates when applied to images containing irregularities. Use of deep learning in this field remains limited due to the lack of publicly available datasets. In this study, as a novel approach, problematic SPECT images —where actual segmentation problems occur- are specifically targeted, and a model capable of segmenting cardiac tissue from such images was developed. The proposed framework leverages ground truth masks generated from easily segmentable images using a shape prior model to train a problem specific designed novel GAN-based model. This GAN was then used to create challenging counterparts of easy cases, which were subsequently used to train a separate deep learning model for robust segmentation including hard to segment images. Compared to existing studies, this work includes a larger number of patient data and provides comparative analyses against the baseline U-Net performance reported in prior literature. Our best performing model not only has the capability of segmenting LV on hard-to-segment images where a commonly used geometric shape and common U-Net models in LV segmentation fails, but also achieves superior performance on easy-to-segment images.</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

Deep Learning-Based Segmentation of Challenging Cardiac SPECT Images: A Robust Approach for Handling Perfusion Defects and Out-of-Myocardium Artifacts

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Efe Arin,
  • I. Samil Yetik,
  • Jovan G. Brankov,
  • Yongyi Yang,
  • Michael A. King

摘要

The function of the heart left ventricle (LV) plays a critical role within the human circulatory system and can be analyzed with SPECT images. Due to the inherent nature of SPECT images being low-resolution, noisy, and low-contrast, myocardial segmentation becomes a challenging task. Manual processes such as tightly cropping the heart region become necessary to partially eliminate part of these issues. For cardiac images without problematic regions, non–AI-based approaches utilizing shape priors on manually cropped images perform sufficiently well; however, their performance significantly deteriorates when applied to images containing irregularities. Use of deep learning in this field remains limited due to the lack of publicly available datasets. In this study, as a novel approach, problematic SPECT images —where actual segmentation problems occur- are specifically targeted, and a model capable of segmenting cardiac tissue from such images was developed. The proposed framework leverages ground truth masks generated from easily segmentable images using a shape prior model to train a problem specific designed novel GAN-based model. This GAN was then used to create challenging counterparts of easy cases, which were subsequently used to train a separate deep learning model for robust segmentation including hard to segment images. Compared to existing studies, this work includes a larger number of patient data and provides comparative analyses against the baseline U-Net performance reported in prior literature. Our best performing model not only has the capability of segmenting LV on hard-to-segment images where a commonly used geometric shape and common U-Net models in LV segmentation fails, but also achieves superior performance on easy-to-segment images.