Margin-Consistent Deep Subtyping of Invasive Lung Adenocarcinoma via Perturbation Fidelity in Whole-Slide Image Analysis

A research team led by Meghdad Sabouri Rad has developed a novel AI framework that dramatically improves the accuracy of classifying invasive lung adenocarcinoma subtypes from whole-slide pathology images. Their approach combines attention-weighted patch aggregation with margin-aware training and introduces a new metric called Perturbation Fidelity (PF) scoring to counteract the over-clustering tendencies of contrastive regularization. The system was trained on the BMIRDS-LUAD dataset containing 203,226 patches from 143 whole-slide images spanning five adenocarcinoma subtypes.

Results show remarkable performance improvements: Vision Transformer-Large achieved 95.20% accuracy (±4.65%), representing a 40% error reduction from the 92.00% baseline, while ResNet101 with attention mechanisms reached 95.89% accuracy (±5.37%), cutting errors by 50% from 91.73%. All five subtypes exceeded an AUC of 0.99, indicating near-perfect discrimination capability. The framework demonstrated strong cross-institutional generalizability, maintaining 80.1% accuracy on the external WSSS4LUAD benchmark despite 15-20% domain-shift-related degradation.

The technical innovation lies in the margin consistency framework that ensures robust feature-logit space alignment, achieving Kendall correlations of 0.88 during training and 0.64 during validation. The PF scoring system imposes structured perturbations through Bayesian-optimized parameters, preserving fine-grained morphological variations that are crucial for accurate subtyping. This represents a significant advancement in making AI pathology tools more reliable in real-world clinical settings where imaging perturbations are common.