Scholarlink

<p>The evaluation of HER2 gene amplification is a time-consuming process that requires quantifying a large number of signals in cancer cells to ensure reproducible results, a task that can be assisted by image analysis (IA) tools. This study aimed to develop and validate an artificial intelligence-based cancer cell detection model for detecting breast carcinoma in bright-field in situ hybridization (ISH) whole-slide images, with the goal of improving cancer cell selection prior to the quantification of HER2 and CEP17 gene copies. The difference in IA performance was not statistically significant compared to visual assessment in quantifying the HER2/CEP17 ratio and average HER2 copy numbers per cell. However, IA with the Cancer Cell Detection Module (IA 2.0) quantified consistently higher values than IA without (IA 1.0). Cancer cell detection in IA 1.0 showed 64% sensitivity, 83% specificity, and a positive predictive value (PPV) above 95% for samples with at least 84% cancer cellularity. IA 2.0 improved cancer cell detection sensitivity to 69% and specificity to 93%, achieving a PPV above 95% for samples with at least 66% cancer cellularity. IA 2.0 showed 95% concordance with visual scoring (<i>k</i> = 0.90), with 90% sensitivity and 100% specificity. Discordant cases required the selection of a second cancer cell-enriched region of interest, shifting classification from HER2-negative to HER2-positive. After correction, IA 2.0 achieved full agreement with visual classification (100%, <i>k</i> = 1.000), underscoring the critical role of the pathologist supervision in this assay.</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

Artificial intelligence enhances cancer cell selection for HER2 gene amplification assessment in breast carcinoma

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月24日

查看全文
  • Iván Rienda,
  • João Pinto,
  • Catarina Eloy,
  • António Polónia

摘要

The evaluation of HER2 gene amplification is a time-consuming process that requires quantifying a large number of signals in cancer cells to ensure reproducible results, a task that can be assisted by image analysis (IA) tools. This study aimed to develop and validate an artificial intelligence-based cancer cell detection model for detecting breast carcinoma in bright-field in situ hybridization (ISH) whole-slide images, with the goal of improving cancer cell selection prior to the quantification of HER2 and CEP17 gene copies. The difference in IA performance was not statistically significant compared to visual assessment in quantifying the HER2/CEP17 ratio and average HER2 copy numbers per cell. However, IA with the Cancer Cell Detection Module (IA 2.0) quantified consistently higher values than IA without (IA 1.0). Cancer cell detection in IA 1.0 showed 64% sensitivity, 83% specificity, and a positive predictive value (PPV) above 95% for samples with at least 84% cancer cellularity. IA 2.0 improved cancer cell detection sensitivity to 69% and specificity to 93%, achieving a PPV above 95% for samples with at least 66% cancer cellularity. IA 2.0 showed 95% concordance with visual scoring (k = 0.90), with 90% sensitivity and 100% specificity. Discordant cases required the selection of a second cancer cell-enriched region of interest, shifting classification from HER2-negative to HER2-positive. After correction, IA 2.0 achieved full agreement with visual classification (100%, k = 1.000), underscoring the critical role of the pathologist supervision in this assay.