Scholarlink

Objectives <p>To assess treatment response in osteosarcoma, two automated convolutional neural networks (CNNs) were developed to quantify tumour volumes and predict response to induction chemotherapy using histopathology as the reference standard.</p> Materials and methods <p>This retrospective, multicentre study included magnetic resonance imaging (MRI) scans from osteosarcoma patients acquired between January 2006 and July 2024. A 3D U-Net CNN segmented tumours and calculated volumes at baseline and post-chemotherapy. A second CNN predicted treatment response based on MRI-derived tumour volume changes using histopathologic necrosis (≥ 90%) as the reference standard. Both models were trained on 162 scans from 81 patients (Centre A) and validated on 40 scans from 20 patients (10 per centre) with Centre B as the external test set. Human readers measured 3D tumour diameters and volumes, compared with CNN-derived volumes using Spearman’s correlation, Bland–Altman plots, and Dice coefficients. Prediction performance was assessed using accuracy, sensitivity, and specificity, with significance determined by agreement metrics.</p> Results <p>Patients from Centre A had a mean age of 15 ± 5 years (52 males), and from Centre B a mean age of 13 ± 0 years (8 males). CNN- and human-derived tumour volumes showed strong correlation (Centre A: <i>r</i> = 0.98, Centre B: <i>r</i> = 0.95; <i>p</i> &lt; 0.001). Dice coefficients were 0.86 (Centre A) and 0.81 (Centre B), with median Hausdorff distances of 15.0 mm and 14.2 mm. The response prediction model classified 16/20 cases (80% accuracy) with 90% sensitivity and 70% specificity.</p> Conclusion <p>CNN-derived tumour volume measurements were comparable to human assessments. CNN-based volume changes predicted histopathologic response to chemotherapy in paediatric osteosarcoma.</p> Key Points <p><Emphasis Type="BoldItalic">Question</Emphasis> <i>Accurate, noninvasive assessment of treatment response in paediatric osteosarcoma is limited by its reliance on manual tumour measurements and post-surgical histopathology.</i></p> <p><Emphasis Type="BoldItalic">Findings</Emphasis> <i>Automated deep learning accurately measured tumour volumes on MRI and predicted chemotherapy response with 80% accuracy, 90% sensitivity, and 70% specificity.</i></p> <p><Emphasis Type="BoldItalic">Clinical relevance</Emphasis> <i>Automated deep learning enables accurate tumour volume assessment and prediction of chemotherapy response in paediatric osteosarcoma, offering a noninvasive tool to support and refine patient management</i>.</p> Graphical Abstract <p></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

Deep learning for accurate tumour volume measurement and prediction of therapy response in paediatric osteosarcoma

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月1日

查看全文
  • Ricarda von Krüchten,
  • Michael Barrow,
  • Lisa Adams,
  • Shashi Bhushan Singh,
  • Zahra Shokri Varniab,
  • Vidyani Suryadevara,
  • Prinska Ghimire,
  • Allison Pribnow,
  • Jing Qi,
  • Dylan Applin,
  • Yashas Ullas Lokesha,
  • Kerem Nernekli,
  • Heike E. Daldrup-Link

摘要

Objectives

To assess treatment response in osteosarcoma, two automated convolutional neural networks (CNNs) were developed to quantify tumour volumes and predict response to induction chemotherapy using histopathology as the reference standard.

Materials and methods

This retrospective, multicentre study included magnetic resonance imaging (MRI) scans from osteosarcoma patients acquired between January 2006 and July 2024. A 3D U-Net CNN segmented tumours and calculated volumes at baseline and post-chemotherapy. A second CNN predicted treatment response based on MRI-derived tumour volume changes using histopathologic necrosis (≥ 90%) as the reference standard. Both models were trained on 162 scans from 81 patients (Centre A) and validated on 40 scans from 20 patients (10 per centre) with Centre B as the external test set. Human readers measured 3D tumour diameters and volumes, compared with CNN-derived volumes using Spearman’s correlation, Bland–Altman plots, and Dice coefficients. Prediction performance was assessed using accuracy, sensitivity, and specificity, with significance determined by agreement metrics.

Results

Patients from Centre A had a mean age of 15 ± 5 years (52 males), and from Centre B a mean age of 13 ± 0 years (8 males). CNN- and human-derived tumour volumes showed strong correlation (Centre A: r = 0.98, Centre B: r = 0.95; p < 0.001). Dice coefficients were 0.86 (Centre A) and 0.81 (Centre B), with median Hausdorff distances of 15.0 mm and 14.2 mm. The response prediction model classified 16/20 cases (80% accuracy) with 90% sensitivity and 70% specificity.

Conclusion

CNN-derived tumour volume measurements were comparable to human assessments. CNN-based volume changes predicted histopathologic response to chemotherapy in paediatric osteosarcoma.

Key Points

Question Accurate, noninvasive assessment of treatment response in paediatric osteosarcoma is limited by its reliance on manual tumour measurements and post-surgical histopathology.

Findings Automated deep learning accurately measured tumour volumes on MRI and predicted chemotherapy response with 80% accuracy, 90% sensitivity, and 70% specificity.

Clinical relevance Automated deep learning enables accurate tumour volume assessment and prediction of chemotherapy response in paediatric osteosarcoma, offering a noninvasive tool to support and refine patient management.

Graphical Abstract