深層学習ベースの仮想エラスチン染色により肺癌の臓側胸膜浸潤評価が向上

Accurate assessment of visceral pleural invasion (VPI) is essential for staging and prognostication in non-small cell lung cancer (NSCLC), yet distinguishing elastin-rich pleural layers on routine hematoxylin and eosin (H&E) sections remains a diagnostic challenge. To overcome the cost and workflow delays associated with special elastic stains, we developed a deep learning pipeline that generates a virtual elastin stain, termed synthetic eosin-based elastin fluorescence (EBEF), directly from standard brightfield H&E slides. A key innovation of this study was the use of intrinsic eosin fluorescence from the same H&E section to create a perfectly co-registered, high-fidelity ground truth for training a conditional generative adversarial network, eliminating the spatial mismatches common in multi-slide approaches. In a multi-institutional validation, supplementing H&E review with synthetic EBEF significantly improved pathologists' diagnostic accuracy for VPI compared with H&E alone (P < 0.0001). Notably, the pre-analytical factors that optimized model performance, including thinner tissue sections (1-3 μm) and high-resolution scanning, also enhanced the perceptual contrast of elastin for pathologists, demonstrating a strong synergy between computational and conventional diagnostic optimization. This study establishes and validates a robust framework for high-fidelity virtual staining that improves diagnostic accuracy and provides a scalable pathway for integrating deep learning-based tools into routine digital pathology. The proposed approach offers a practical and cost-effective alternative to ancillary special stains in NSCLC evaluation.