Scholarlink

Objective <p>Diagnosing and grading chondroid bone tumors with radiography is difficult due to overlapping features with benign conditions such as avascular necrosis and fibrous dysplasia. Histology remains the gold standard, but it is invasive, costly, and not always accessible. This pilot study introduces a contrastive learning framework to address two challenges: (1) developing an AI system for accurate tumor classification and grading using radiographs alone, and (2) creating an enhanced multimodal pipeline when histology is available.</p> Materials and methods <p>We retrospectively analyzed radiographs from 188 patients and histology images from 63 patients at a tertiary academic medical center in the United States. The stepwise framework included: (1) unimodal classification, (2) contrastive learning to align cross-modal embeddings, and (3) classification using enhanced representations. Models were trained with 5-fold cross-validation and evaluated using AUC, accuracy, sensitivity, and specificity.</p> Results <p>Our framework demonstrates promising performance across all tasks. The radiograph-based model achieved an AUC of 0.91 (95%CI: 0.82–1.00) in distinguishing tumors from avascular necrosis and fibrous dysplasia. For grading, contrastive learning improved radiograph-only performance from AUC 0.86 to 0.95 (95%CI: 0.85–1.00). The histology-only model improved from AUC 0.73 to 0.83 with contrastive enhancement. Multimodal integration achieved perfect discrimination (AUC = 1.00) on the available subset.</p> Conclusion <p>This study establishes proof-of-concept that contrastive learning can effectively bridge radiographic and histological representations for tumor assessment. The framework offers clinical potential by enabling non-invasive classification in resource-limited settings while allowing multimodal enhancement. These results warrant validation in larger, multi-institutional cohorts.</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

Multimodal contrastive learning for non-invasive chondroid bone tumor classification and grading using radiographs

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月20日

查看全文
  • Peiying Hua,
  • Jessica M. Sin,
  • Eric R. Henderson,
  • Jason Ha,
  • Darcy A. Kerr,
  • Saeed Hassanpour

摘要

Objective

Diagnosing and grading chondroid bone tumors with radiography is difficult due to overlapping features with benign conditions such as avascular necrosis and fibrous dysplasia. Histology remains the gold standard, but it is invasive, costly, and not always accessible. This pilot study introduces a contrastive learning framework to address two challenges: (1) developing an AI system for accurate tumor classification and grading using radiographs alone, and (2) creating an enhanced multimodal pipeline when histology is available.

Materials and methods

We retrospectively analyzed radiographs from 188 patients and histology images from 63 patients at a tertiary academic medical center in the United States. The stepwise framework included: (1) unimodal classification, (2) contrastive learning to align cross-modal embeddings, and (3) classification using enhanced representations. Models were trained with 5-fold cross-validation and evaluated using AUC, accuracy, sensitivity, and specificity.

Results

Our framework demonstrates promising performance across all tasks. The radiograph-based model achieved an AUC of 0.91 (95%CI: 0.82–1.00) in distinguishing tumors from avascular necrosis and fibrous dysplasia. For grading, contrastive learning improved radiograph-only performance from AUC 0.86 to 0.95 (95%CI: 0.85–1.00). The histology-only model improved from AUC 0.73 to 0.83 with contrastive enhancement. Multimodal integration achieved perfect discrimination (AUC = 1.00) on the available subset.

Conclusion

This study establishes proof-of-concept that contrastive learning can effectively bridge radiographic and histological representations for tumor assessment. The framework offers clinical potential by enabling non-invasive classification in resource-limited settings while allowing multimodal enhancement. These results warrant validation in larger, multi-institutional cohorts.