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

TL;DR

This paper introduces a margin consistency framework with perturbation fidelity for robust whole-slide image classification of lung adenocarcinoma subtypes, significantly improving accuracy and generalizability across datasets.

Contribution

It proposes a novel margin consistency approach combined with structured perturbations and attention mechanisms, enhancing robustness and cross-institutional performance in lung cancer subtyping.

Findings

Achieved 95.20% accuracy with Vision Transformer-Large, reducing error by 40%.

ResNet101 with attention reached 95.89% accuracy, 50% error reduction.

All subtypes exceeded 0.99 AUC, demonstrating high classification performance.

Abstract

Whole-slide image classification for invasive lung adenocarcinoma subtyping remains vulnerable to real-world imaging perturbations that undermine model reliability at the decision boundary. We propose a margin consistency framework evaluated on 203,226 patches from 143 whole-slide images spanning five adenocarcinoma subtypes in the BMIRDS-LUAD dataset. By combining attention-weighted patch aggregation with margin-aware training, our approach achieves robust feature-logit space alignment measured by Kendall correlations of 0.88 during training and 0.64 during validation. Contrastive regularization, while effective at improving class separation, tends to over-cluster features and suppress fine-grained morphological variation; to counteract this, we introduce Perturbation Fidelity (PF) scoring, which imposes structured perturbations through Bayesian-optimized parameters. Vision Transformer-Large achieves 95.20 +/- 4.65% accuracy, representing a 40% error reduction from the 92.00 +/- 5.36% baseline, while ResNet101 with an attention mechanism reaches 95.89 +/- 5.37% from 91.73 +/- 9.23%, a 50% error reduction. All five subtypes exceed an area under the receiver operating characteristic curve (AUC) of 0.99. On the WSSS4LUAD external benchmark, ResNet50 with an attention mechanism attains 80.1% accuracy, demonstrating cross-institutional generalizability despite approximately 15-20% domain-shift-related degradation and identifying opportunities for future adaptation research.