FourierMIL: Fourier Filtering-based Multiple Instance Learning for Whole Slide Image Analysis

Area of Science:

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

Background:

• Computer vision techniques like CNNs and transformers advance image classification.
• Gigapixel whole-slide images (WSIs) in digital pathology pose challenges due to size and heterogeneity.
• Existing methods struggle with the scale and complexity of WSIs.

Purpose of the Study:

• Introduce FourierMIL, a multiple instance learning framework for efficient WSI analysis.
• Leverage the discrete Fourier transform to capture global and local dependencies in WSIs.
• Demonstrate FourierMIL's adaptability across diverse digital stains and pathology tasks.

Main Methods:

• Developed FourierMIL, an attention-free multiple instance learning framework.
• Utilized the discrete Fourier transform for feature extraction from WSIs.
• Evaluated FourierMIL on metastasis detection (CAMELYON16), lung cancer classification (TCGA, CPTAC), and Alzheimer's disease pathology identification (UNITE, FHS, ADC).

Main Results:

• FourierMIL achieved superior performance across all tested digital pathology tasks.
• Demonstrated robustness in metastasis detection on H&E-stained lymph node WSIs.
• Showcased effectiveness in lung cancer classification and Alzheimer's disease pathology identification on diverse datasets.

Conclusions:

• FourierMIL offers a versatile and robust solution for digital pathology.
• The framework efficiently handles large-scale WSIs, overcoming limitations of conventional methods.
• FourierMIL's attention-free approach provides a scalable alternative for various pathology applications.