Robust Histopathology Subtyping via Perturbation Fidelity in Deep Classifier.

Area of Science:

• Computational pathology
• Medical imaging analysis
• Artificial intelligence in oncology

Background:

• Deep learning models for invasive lung adenocarcinoma subtyping are susceptible to real-world imaging variations.
• Accurate subtyping is crucial for effective lung cancer treatment and prognosis.

Purpose of the Study:

• To develop a robust deep learning framework for invasive lung adenocarcinoma subtyping resistant to imaging perturbations.
• To enhance the accuracy and reliability of automated subtyping using whole-slide images.

Main Methods:

• Implemented a margin consistency framework integrating attention-weighted aggregation and margin-aware training.
• Introduced Perturbation Fidelity scoring with Bayesian-optimized parameters to mitigate feature over-clustering.
• Evaluated Vision Transformer-Large and ResNet101 models on the BMIRDS-LUAD dataset.

Main Results:

• Achieved significant error reduction in subtyping accuracy for both Vision Transformer-Large (40%) and ResNet101 (50%).
• Demonstrated strong feature-logit space alignment with high Kendall correlations (0.88 training, 0.64 validation).
• Attained excellent performance across all five subtypes with area under ROC curves exceeding 0.99.

Conclusions:

• The proposed margin consistency framework enhances the robustness and accuracy of deep learning for lung adenocarcinoma subtyping.
• The method shows promise for clinical application, though domain adaptation research is needed to address performance variations across institutions.