Re-identification from histopathology images

Area of Science:

• Digital pathology
• Medical imaging analysis
• Machine learning in healthcare

Background:

• Deep learning models analyze histopathology images for tasks like tumor subtyping.
• Large datasets are crucial for training these models but require anonymization to protect patient privacy.
• Existing anonymization methods may not fully mitigate re-identification risks.

Purpose of the Study:

• To assess the re-identification risk of patients in histopathology datasets using deep learning.
• To compare the performance of various deep learning models for patient re-identification.
• To develop a risk assessment scheme for privacy protection in digital pathology.

Main Methods:

• Evaluation of simple and state-of-the-art deep learning algorithms for patient re-identification.
• Utilized two TCIA datasets (lung squamous cell carcinoma, lung adenocarcinoma) and an in-house meningioma dataset.
• Compared whole slide image classifiers and feature extractors for re-identification accuracy.

Main Results:

• Deep learning algorithms achieved substantial accuracy in re-identifying patients from histopathology slides.
• Achieved F1 scores up to 80.1% on LSCC, 77.19% on LUAD, and 77.09% on meningioma datasets.
• Demonstrated that even simpler models can pose significant re-identification risks.

Conclusions:

• Deep learning poses a considerable risk of patient re-identification in histopathology datasets.
• A risk assessment scheme is proposed to evaluate and mitigate privacy risks before data publication.
• Highlights the need for robust anonymization techniques in digital pathology.