Scholarlink

Background <p>Bone marrow (BM) is the most common site of metastatic disease at diagnosis and a frequent site of relapse in neuroblastoma. Digital cytology images of BM smears offer a rich data source for artificial intelligence models, which may potentially facilitate more cost-effective risk stratification within the diagnostic workflow. This study aims to develop an interpretable cytology model for detecting BM metastasis in pediatric neuroblastoma.</p> Methods <p>This retrospective diagnostic study used Wright-Giemsa–stained BM cytology images from 359 neuroblastoma patients who underwent BM screening between January 2019 and June 2024 across multiple centers in China. After the quality evaluation, we generated 1384,007 patches from BM digital cytology to develop and validate the cytology model. In the model construction, we integrated a multiple-instance learning framework with convolutional neural networks to extract cytology features, referred to as cMIL. The cytology model was trained for BM metastasis detection and risk stratification with interpretability.</p> Results <p>For metastasis detection, the cytology model achieved an AUC of 0.924 (95% CI, 0.775–1.000) in the training cohort. Performance remained strong in external validation, with AUCs of 0.826 (95% CI, 0.741–0.911) in Cohort A and 0.795 (95% CI, 0.684–0.906) in Cohort B, indicating consistent performance across independent multicenter cohorts. The cMIL score also successfully stratified patients in terms of survival outcomes (log-rank <i>p</i> &lt; 0.05). Interpretability analyses further demonstrated that the model’s predictions were associated with clinically relevant cytological features.</p> Conclusions <p>In this retrospective diagnostic study, the developed cytology model demonstrated high discriminative performance in detecting BM metastasis and captured the underlying complexity and heterogeneity of BM. These findings suggest that the cytology model could serve as a promising tool for improving metastasis detection and risk stratification in patients with neuroblastoma, potentially contributing to personalized treatment strategies and enhanced disease monitoring.</p> Trial registration <p>This retrospective study was registered with ClinicalTrials.gov (NCT06703944) on November 21, 2024. Study title: bone marrow cytology-based artificial intelligence model for detection and prognosis of neuroblastoma. (<a href="https://register.clinicaltrials.gov">https://register.clinicaltrials.gov</a>).</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

Interpretative diagnostic model for neuroblastoma metastases using bone marrow cytology

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月7日

查看全文
  • Juan Ma,
  • Qiang Yao,
  • Xiaoying Fu,
  • Zhouqi Xia,
  • Yaoting Yue,
  • Dongqing Xu,
  • Xiaojun Yuan,
  • Liebin Zhao,
  • Jinhu Wang,
  • Ao Dong,
  • Limei Gao,
  • Junyao Yang,
  • Lisong Shen,
  • Yingxia Zheng,
  • Shaoqing Ni

摘要

Background

Bone marrow (BM) is the most common site of metastatic disease at diagnosis and a frequent site of relapse in neuroblastoma. Digital cytology images of BM smears offer a rich data source for artificial intelligence models, which may potentially facilitate more cost-effective risk stratification within the diagnostic workflow. This study aims to develop an interpretable cytology model for detecting BM metastasis in pediatric neuroblastoma.

Methods

This retrospective diagnostic study used Wright-Giemsa–stained BM cytology images from 359 neuroblastoma patients who underwent BM screening between January 2019 and June 2024 across multiple centers in China. After the quality evaluation, we generated 1384,007 patches from BM digital cytology to develop and validate the cytology model. In the model construction, we integrated a multiple-instance learning framework with convolutional neural networks to extract cytology features, referred to as cMIL. The cytology model was trained for BM metastasis detection and risk stratification with interpretability.

Results

For metastasis detection, the cytology model achieved an AUC of 0.924 (95% CI, 0.775–1.000) in the training cohort. Performance remained strong in external validation, with AUCs of 0.826 (95% CI, 0.741–0.911) in Cohort A and 0.795 (95% CI, 0.684–0.906) in Cohort B, indicating consistent performance across independent multicenter cohorts. The cMIL score also successfully stratified patients in terms of survival outcomes (log-rank p < 0.05). Interpretability analyses further demonstrated that the model’s predictions were associated with clinically relevant cytological features.

Conclusions

In this retrospective diagnostic study, the developed cytology model demonstrated high discriminative performance in detecting BM metastasis and captured the underlying complexity and heterogeneity of BM. These findings suggest that the cytology model could serve as a promising tool for improving metastasis detection and risk stratification in patients with neuroblastoma, potentially contributing to personalized treatment strategies and enhanced disease monitoring.

Trial registration

This retrospective study was registered with ClinicalTrials.gov (NCT06703944) on November 21, 2024. Study title: bone marrow cytology-based artificial intelligence model for detection and prognosis of neuroblastoma. (https://register.clinicaltrials.gov).