Scholarlink

<p>Accurate classification of pulmonary nodule subtypes is vital for early lung cancer management. Yet, building reliable AI models remains challenging due to limited annotated data and the need to jointly exploit the complementary spatial cues offered by 2D CT slices and 3D volumetric scans. This work introduces a multimodal meta-learning framework for few-shot lung nodule classification, designed to enhance data-efficiency, fusion effectiveness, and interpretability. A contrastive 2D–3D cross-modal alignment module first harmonises slice-level and volume-level encoders into a unified latent space, enabling stable multimodal fusion under scarce supervision. A Prototypical Network-based episodic learner is then employed to support rapid adaptation to unseen nodules with only a few labelled samples. Beyond accuracy, the framework incorporates prototype-guided saliency mapping, fusion-margin analysis, and cognitive consistency metrics to assess the trustworthiness of model decisions. Experiments on a rare-data clinical cohort demonstrated that the proposed approach achieved 85.1% accuracy (95% CI [0.825, 0.877]) in a five-shot setting, surpassing single-modality and supervised baselines by 6–8%. Performance gains were statistically significant under the paired Wilcoxon signed-rank test (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p &lt; 0.05\)</EquationSource> </InlineEquation>). The results demonstrate that cross-modal alignment with few-shot meta-learning delivers a cognitively robust, transparent, and data-efficient solution for multimodal lung nodule analysis, with potential to support triage in low-resource lung cancer screening programs. The codebase implementing the proposed multimodal meta-learning and calibration framework is publicly accessible for reproducibility at <a href="https://github.com/JuhiGupta2023/multimodal-meta-lung">https://github.com/JuhiGupta2023/multimodal-meta-lung</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

Multimodal meta-learning for lung nodule classification under few-shot settings: a trustworthy AI framework with 2D–3D cross-modal alignment

  • 期刊论文

  • 发表日期：

    2026年2月17日

查看全文
  • Juhi Gupta,
  • Monica Mehrotra,
  • Arpita Aggarwal

摘要

Accurate classification of pulmonary nodule subtypes is vital for early lung cancer management. Yet, building reliable AI models remains challenging due to limited annotated data and the need to jointly exploit the complementary spatial cues offered by 2D CT slices and 3D volumetric scans. This work introduces a multimodal meta-learning framework for few-shot lung nodule classification, designed to enhance data-efficiency, fusion effectiveness, and interpretability. A contrastive 2D–3D cross-modal alignment module first harmonises slice-level and volume-level encoders into a unified latent space, enabling stable multimodal fusion under scarce supervision. A Prototypical Network-based episodic learner is then employed to support rapid adaptation to unseen nodules with only a few labelled samples. Beyond accuracy, the framework incorporates prototype-guided saliency mapping, fusion-margin analysis, and cognitive consistency metrics to assess the trustworthiness of model decisions. Experiments on a rare-data clinical cohort demonstrated that the proposed approach achieved 85.1% accuracy (95% CI [0.825, 0.877]) in a five-shot setting, surpassing single-modality and supervised baselines by 6–8%. Performance gains were statistically significant under the paired Wilcoxon signed-rank test ( \(p < 0.05\) ). The results demonstrate that cross-modal alignment with few-shot meta-learning delivers a cognitively robust, transparent, and data-efficient solution for multimodal lung nodule analysis, with potential to support triage in low-resource lung cancer screening programs. The codebase implementing the proposed multimodal meta-learning and calibration framework is publicly accessible for reproducibility at https://github.com/JuhiGupta2023/multimodal-meta-lung.