Fourier Transform Multiple Instance Learning for whole slide image classification.

Area of Science:

• Computational pathology
• Digital pathology
• Medical image analysis

Background:

• Whole slide image (WSI) classification commonly uses multiple instance learning (MIL) with spatial patch features.
• Current MIL methods face challenges in capturing global dependencies due to WSI size and local patch embeddings, limiting coarse structure modeling for diagnostics.

Purpose of the Study:

• To introduce Fourier Transform Multiple Instance Learning (FFT-MIL), a novel framework designed to integrate global context into WSI classification.
• To address the limitations of existing MIL approaches in modeling coarse structures by incorporating frequency-domain information.

Main Methods:

• FFT-MIL augments standard MIL with a frequency-domain branch using Fast Fourier Transform (FFT) to extract low-frequency crops from WSIs.
• A modular FFT-Block, featuring convolutional layers and Min-Max normalization, processes these frequency crops to generate compact global context.
• The learned global frequency features are fused with spatial patch features via lightweight integration strategies compatible with various MIL architectures.

Main Results:

• FFT-MIL was evaluated by integrating the FFT-Block into six state-of-the-art MIL methods across three public datasets (BRACS, LUAD, IMP).
• The integration consistently improved macro F1 scores by an average of 3.51% and area under the curve by 1.51% across different architectures and datasets.

Conclusions:

• FFT-MIL demonstrates the efficacy of frequency-domain learning for capturing global dependencies in WSI classification.
• This approach complements spatial features, enhancing the scalability and accuracy of MIL-based computational pathology.
• The study provides a publicly available codebase for FFT-MIL, promoting further research and application.