Scholarlink

<p>Postoperative breast radiotherapy after Breast-Conserving surgery (BCS) has been proven effective in treating breast cancers. Cone Beam Computed Tomography (CBCT) guided radiation therapy (RT) facilitates appropriate patient positioning by mapping the target contours on planning Computed Tomography (pCT) to CBCT images for executing accurate dose delivery. However, the on-board CBCT images are of limited quality and possess inconsistent Hounsfield Units (HU). Several AI-enabled methodologies have been proposed to enhance the quality of CBCT images or generate high-quality synthetic Computed Tomography images using generative AI models. However, these methods usually employ unpaired training and fail to track precise anatomical changes, which are paramount in RT. Moreover, the impacts of image normalisation techniques, which standardise the inter-scan variability of HU values to a consistent scale, are not properly explored. Image normalization ensures consistency in image intensity, contrast, and scale, improving the performance of AI models by reducing variability across different imaging systems. To overcome these bottlenecks the proposed work focuses on (i) utilization of different image normalization techniques to standardise the inter-scan variability in the captured CBCT, pCT samples (ii) preparation of paired BCS patient datasets using histogram matching (iii) utilization of U-Net model to enhance the quality of on-board CBCT images (iv) impact analysis of normalization techniques on the performance of U-Net model through extensive quantitative evaluation and visual representation. Furthermore, Explainability tools like LIME and XRAI are used to interpret the effect of normalization techniques by identifying the regions that contribute the most to the image enhancement process. Training a U-Net model on paired datasets minimises the residual mismatching problem and enhances the generalisation capability. Min–Max and Z-Score normalization techniques enable to preserve superior structural and feature similarity in the enhancement process.</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

Impact of Normalization Techniques in CBCT Image Quality Enhancement for Breast-Conserving Surgery Patients

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Papia Banerjee,
  • Kasturi Barik,
  • Rajashree Nayak

摘要

Postoperative breast radiotherapy after Breast-Conserving surgery (BCS) has been proven effective in treating breast cancers. Cone Beam Computed Tomography (CBCT) guided radiation therapy (RT) facilitates appropriate patient positioning by mapping the target contours on planning Computed Tomography (pCT) to CBCT images for executing accurate dose delivery. However, the on-board CBCT images are of limited quality and possess inconsistent Hounsfield Units (HU). Several AI-enabled methodologies have been proposed to enhance the quality of CBCT images or generate high-quality synthetic Computed Tomography images using generative AI models. However, these methods usually employ unpaired training and fail to track precise anatomical changes, which are paramount in RT. Moreover, the impacts of image normalisation techniques, which standardise the inter-scan variability of HU values to a consistent scale, are not properly explored. Image normalization ensures consistency in image intensity, contrast, and scale, improving the performance of AI models by reducing variability across different imaging systems. To overcome these bottlenecks the proposed work focuses on (i) utilization of different image normalization techniques to standardise the inter-scan variability in the captured CBCT, pCT samples (ii) preparation of paired BCS patient datasets using histogram matching (iii) utilization of U-Net model to enhance the quality of on-board CBCT images (iv) impact analysis of normalization techniques on the performance of U-Net model through extensive quantitative evaluation and visual representation. Furthermore, Explainability tools like LIME and XRAI are used to interpret the effect of normalization techniques by identifying the regions that contribute the most to the image enhancement process. Training a U-Net model on paired datasets minimises the residual mismatching problem and enhances the generalisation capability. Min–Max and Z-Score normalization techniques enable to preserve superior structural and feature similarity in the enhancement process.