Scholarlink

<p>Accurate post-lumpectomy target delineation is computationally demanding and challenged by limited annotated data, motivating scalable deep learning solutions for radiotherapy planning. We proposed a hybrid ensemble transfer learning framework based on U-Net architectures with VGG16 and MobileNetV2 backbones. Using 2D axial CT slices extracted from clinical 3D volumes collected at Imam Reza Hospital, Kermanshah, a deep convolutional generative adversarial network (DCGAN) was employed for patient-wise data augmentation to address data scarcity. The GPU-accelerated framework was evaluated in terms of computational performance, scalability, and near-real-time feasibility. The ensemble achieved a mean Intersection over Union (IoU) of 0.94 on test set, outperforming single-model baselines, with inference times of a few seconds per scan on an NVIDIA T4 GPU. Clinician-reviewed evaluations indicated improved contour consistency and reduced delineation uncertainty, supporting more reliable radiotherapy planning. The proposed framework provides an accurate and computationally efficient solution suitable for integration into modern radiotherapy workflows.</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

A scalable ensemble transfer learning framework for post-lumpectomy target segmentation in breast radiotherapy CT images

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Hassan Salarabadi,
  • Dariush Salimi,
  • Negin Farshchian

摘要

Accurate post-lumpectomy target delineation is computationally demanding and challenged by limited annotated data, motivating scalable deep learning solutions for radiotherapy planning. We proposed a hybrid ensemble transfer learning framework based on U-Net architectures with VGG16 and MobileNetV2 backbones. Using 2D axial CT slices extracted from clinical 3D volumes collected at Imam Reza Hospital, Kermanshah, a deep convolutional generative adversarial network (DCGAN) was employed for patient-wise data augmentation to address data scarcity. The GPU-accelerated framework was evaluated in terms of computational performance, scalability, and near-real-time feasibility. The ensemble achieved a mean Intersection over Union (IoU) of 0.94 on test set, outperforming single-model baselines, with inference times of a few seconds per scan on an NVIDIA T4 GPU. Clinician-reviewed evaluations indicated improved contour consistency and reduced delineation uncertainty, supporting more reliable radiotherapy planning. The proposed framework provides an accurate and computationally efficient solution suitable for integration into modern radiotherapy workflows.