Scholarlink

<p>The differential diagnosis between reactive follicular hyperplasia (RFH) and follicular lymphoma (FL) in head and neck tissues represents a diagnostic challenge, particularly given the rarity of extranodal FL at these anatomical sites. We developed and evaluated an explainable multimodal artificial intelligence framework to assist in this differential diagnosis. This international multicentre study assembled 108 cases (54 RFH, 54 FL) from 10 centres across 4 continents, representing one of the largest collections of head and neck follicular lesions. We developed a multimodal framework integrating convolutional neural networks (CNNs: AlexNet, VGG16, ResNet18), vision transformers (CellViT + +), graph neural networks (Cell-GNN for spatial analysis), and gradient-boosted decision trees (XGBoost) with morphometric features. Explainability was achieved through Grad-CAM visualisation and SHAP analysis. The best-performing model (CellViT + +) achieved 95.7% accuracy on the internal test set. External validation demonstrated accuracy of 80.5% (ResNet18, Cohort 1) and 69.0% (VGG16-Seg, Cohort 2), with performance variation reflecting the challenge of generalisation across centres. Explainability analysis revealed that the multimodal framework integrated morphometric features (nuclear area, eccentricity) with epidemiological context (patient age, consistent with known FL demographics). Cell-GNN spatial analysis identified significant architectural reorganisation in FL, quantified by Hedges' g = -6.65, representing the loss of normal follicular polarity. This study demonstrates the feasibility of a multimodal explainable AI framework for assisting in the differential diagnosis of rare head and neck follicular lesions. The integration of morphological, spatial, and clinical features provides a foundation for future collaborative validation studies and potential diagnostic support in specialised pathology settings.</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 multimodal explainable ai framework to assist in the differential diagnosis of head and neck reactive follicular hyperplasia and follicular lymphoma: an international multicentre study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月2日

查看全文
  • Lucas Lacerda de Souza,
  • Zhiyang Chen,
  • Cinthia Veronica Bardález López de Cáceres,
  • Nathalia Gomes Rodrigues,
  • Hélder Antônio Rebelo Pontes,
  • Ciro Dantas Soares,
  • Fábio Luiz Coracin,
  • Igor Kimura,
  • Sheila Aparecida Coelho Siqueira,
  • Fernanda Viviane Mariano,
  • Talita de Carvalho Kimura,
  • João Felipe Leite Bonfitto,
  • Leandro Luiz Lopes de Freitas,
  • Sibele Nascimento de Aquino,
  • Paul Hankinson,
  • Hanya Mahmood,
  • Hannah Walsh,
  • Itzia Araceli Torres Torres,
  • Adalberto Mosqueda-Taylor,
  • Pedro Bandeira Aleixo,
  • Manoela Domingues Martins,
  • Carolina Louzada Menna-Barreto,
  • Ahmed Hagag,
  • Thaís Cerqueira Reis Nakamura,
  • Giovanna Calabrese dos Santos,
  • Matheus Cardoso Moraes,
  • Marcio Ajudarte Lopes,
  • Alan Roger Santos-Silva,
  • Felipe Paiva Fonseca,
  • Syed Ali Khurram,
  • Pablo Agustin Vargas

摘要

The differential diagnosis between reactive follicular hyperplasia (RFH) and follicular lymphoma (FL) in head and neck tissues represents a diagnostic challenge, particularly given the rarity of extranodal FL at these anatomical sites. We developed and evaluated an explainable multimodal artificial intelligence framework to assist in this differential diagnosis. This international multicentre study assembled 108 cases (54 RFH, 54 FL) from 10 centres across 4 continents, representing one of the largest collections of head and neck follicular lesions. We developed a multimodal framework integrating convolutional neural networks (CNNs: AlexNet, VGG16, ResNet18), vision transformers (CellViT + +), graph neural networks (Cell-GNN for spatial analysis), and gradient-boosted decision trees (XGBoost) with morphometric features. Explainability was achieved through Grad-CAM visualisation and SHAP analysis. The best-performing model (CellViT + +) achieved 95.7% accuracy on the internal test set. External validation demonstrated accuracy of 80.5% (ResNet18, Cohort 1) and 69.0% (VGG16-Seg, Cohort 2), with performance variation reflecting the challenge of generalisation across centres. Explainability analysis revealed that the multimodal framework integrated morphometric features (nuclear area, eccentricity) with epidemiological context (patient age, consistent with known FL demographics). Cell-GNN spatial analysis identified significant architectural reorganisation in FL, quantified by Hedges' g = -6.65, representing the loss of normal follicular polarity. This study demonstrates the feasibility of a multimodal explainable AI framework for assisting in the differential diagnosis of rare head and neck follicular lesions. The integration of morphological, spatial, and clinical features provides a foundation for future collaborative validation studies and potential diagnostic support in specialised pathology settings.