Deceptive learning in histopathology

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Histopathology image analysis

Background:

• Deep learning (DL) offers potential for automating histopathology tasks and discovering novel biological insights.
• Systematic evaluation of the trustworthiness of visual strategies learned by DL models in histopathology is lacking.

Purpose of the Study:

• To systematically evaluate deep neural networks (DNNs) trained for histopathological analysis.
• To determine if DNNs' learned strategies are trustworthy or deceptive.

Main Methods:

• Trained various DNNs on 221 whole-slide images (WSIs) of lung adenocarcinoma.
• Evaluated DNNs on molecular profiling (KRAS vs. EGFR mutations), primary tissue determination, and tumor detection.

Main Results:

• DNNs achieved above-chance performance in molecular profiling by exploiting correlations between histological subtypes and mutations, failing to generalize to laser capture microdissection (LCM) test sets.
• DNNs learned robust and trustworthy strategies for primary tissue determination and tumor detection/localization.

Conclusions:

• DNNs show promise for aiding pathologists but can learn deceptive strategies using spurious correlations, rendering them unsuitable for research or clinical use.
• A proposed framework for model evaluation and interpretation is crucial for developing reliable automated histopathological analysis systems.