Scholarlink

<p>Early and reliable identification of malignant lung nodules remains a persistent clinical challenge due to scanner variability, heterogeneous nodule appearance, and the limited calibration of existing deep learning models. Although Vision Transformers (ViTs) have demonstrated strong global contextual modeling, single-backbone architectures often struggle with generalization and uncertainty in real-world conditions. To address these limitations, we propose an adaptive Mixture-of-Experts (MoE) ensemble that integrates four state-of-the-art ViT-family backbones Swin Transformer, Vision Transformer (ViT), BEiT, and EVA within a 3D nodule-centric diagnostic framework. A lightweight gating network assigns case-specific expert weights using radiomics descriptors and acquisition-level features, enabling dynamic expert selection tailored to individual nodule characteristics. Further, we introduce a cross-attention fusion module for combining radiomics with transformer embeddings, Monte-Carlo dropout–based uncertainty estimation, and post-hoc temperature scaling for probability calibration. Comprehensive evaluation on LIDC-IDRI with external validation on LUNA16 demonstrates that the proposed MoE achieves superior discrimination (AUC), higher malignant recall, more reliable calibration, and improved clinical net benefit compared with single transformer models and static ensembles. Counterfactual visualizations and token-level explanations further enhance transparency and clinical interpretability. The proposed framework offers a robust, calibrated, and uncertainty-aware AI system for early lung cancer diagnosis and represents a step toward trustworthy real-world deployment.</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

Adaptive Mixture-of-Experts Vision Transformer Ensemble with 3D Nodule Reasoning and Clinical Calibration for Lung Cancer Diagnosis

  • 期刊论文

  • 发表日期：

    2026年5月30日

查看全文
  • Nasareenbanu Devihosur,
  • M. G. Ravi Kumar

摘要

Early and reliable identification of malignant lung nodules remains a persistent clinical challenge due to scanner variability, heterogeneous nodule appearance, and the limited calibration of existing deep learning models. Although Vision Transformers (ViTs) have demonstrated strong global contextual modeling, single-backbone architectures often struggle with generalization and uncertainty in real-world conditions. To address these limitations, we propose an adaptive Mixture-of-Experts (MoE) ensemble that integrates four state-of-the-art ViT-family backbones Swin Transformer, Vision Transformer (ViT), BEiT, and EVA within a 3D nodule-centric diagnostic framework. A lightweight gating network assigns case-specific expert weights using radiomics descriptors and acquisition-level features, enabling dynamic expert selection tailored to individual nodule characteristics. Further, we introduce a cross-attention fusion module for combining radiomics with transformer embeddings, Monte-Carlo dropout–based uncertainty estimation, and post-hoc temperature scaling for probability calibration. Comprehensive evaluation on LIDC-IDRI with external validation on LUNA16 demonstrates that the proposed MoE achieves superior discrimination (AUC), higher malignant recall, more reliable calibration, and improved clinical net benefit compared with single transformer models and static ensembles. Counterfactual visualizations and token-level explanations further enhance transparency and clinical interpretability. The proposed framework offers a robust, calibrated, and uncertainty-aware AI system for early lung cancer diagnosis and represents a step toward trustworthy real-world deployment.