Breast cancer histopathological image classification based on collaborative multi-domain feature learning.

Area of Science:

• Oncology
• Medical Imaging
• Computational Pathology

Background:

• Breast cancer classification is crucial for treatment decisions due to its heterogeneity.
• Current methods like Convolutional Neural Networks (CNNs) and Transformers have limitations in capturing complex pathological features.
• CNNs struggle with irregular cell morphology, while Transformers lack detailed 2D spatial characterization.

Purpose of the Study:

• To introduce a novel Spatial-Frequency Domain Feature Extraction Model (S-FDFEM) for improved breast cancer pathological image recognition.
• To integrate spatial and frequency domain information to overcome limitations of existing deep learning models.
• To enhance the feature learning capabilities for more accurate breast cancer classification.

Main Methods:

• Developed S-FDFEM integrating spatial and frequency domain feature extraction.
• Employed Deformable Bottleneck Convolution (DBottConv) in the spatial domain to capture intricate cell morphology.
• Utilized wavelet low frequencies and Fourier high frequencies in the frequency domain, processed by statistical transformer and depth gradient modules.

Main Results:

• The S-FDFEM demonstrated superior performance in breast cancer pathological image classification.
• Validation on BreakHis and BACH datasets confirmed the model's effectiveness.
• The integrated approach successfully enhanced feature learning for pathological image recognition.

Conclusions:

• The S-FDFEM offers a significant advancement in breast cancer pathological image analysis.
• Integrating spatial and frequency domain features provides a more comprehensive understanding of pathological images.
• This model holds promise for improving clinical diagnosis and treatment decision-making for breast cancer.