Scholarlink

The findings of this review provide an analysis of commonly used deep learning (DL) methods for lung tumor segmentation from computed tomography (CT) images. We examine prominent architectures—such as Convolutional Neural Networks (CNNs) (U-Net, nnUNet), Transformers (TransUNet), and Implicit Neural Representations (INRs) (IOSNet, I-MedSAM)—and compare their efficacy using clinical metrics. Results indicate that hybrid models often outperform conventional designs. This review’s original contribution lies in identifying future research frontiers beyond standard models, including the critical needs for uncertainty quantification (UQ), spatiotemporal analysis for longitudinal monitoring, explainable AI (XAI), and privacy-preserving federated learning. We conclude that the next generation of deployable tools will depend on hybrid, trustworthy, and data-efficient architectures that integrate these capabilities. These findings will be of interest to clinicians, model developers, and researchers in the field.

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 for Lung Tumor Segmentation in Computed Tomography: A Review of Architectures and Future Frontiers

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ferenc Moisi,
  • Laszlo Barna Iantovics

摘要

The findings of this review provide an analysis of commonly used deep learning (DL) methods for lung tumor segmentation from computed tomography (CT) images. We examine prominent architectures—such as Convolutional Neural Networks (CNNs) (U-Net, nnUNet), Transformers (TransUNet), and Implicit Neural Representations (INRs) (IOSNet, I-MedSAM)—and compare their efficacy using clinical metrics. Results indicate that hybrid models often outperform conventional designs. This review’s original contribution lies in identifying future research frontiers beyond standard models, including the critical needs for uncertainty quantification (UQ), spatiotemporal analysis for longitudinal monitoring, explainable AI (XAI), and privacy-preserving federated learning. We conclude that the next generation of deployable tools will depend on hybrid, trustworthy, and data-efficient architectures that integrate these capabilities. These findings will be of interest to clinicians, model developers, and researchers in the field.