Deep Multi-Magnification Similarity Learning for Histopathological Image Classification.

Area of Science:

• Digital Pathology
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Precise histopathological image classification is vital for computer-aided diagnosis.
• Magnification-based learning networks enhance histopathological classification performance.
• Fusing multi-magnification histopathological images remains an under-explored area.

Purpose of the Study:

• To propose a novel deep multi-magnification similarity learning (DSML) approach.
• To enable interpretation of multi-magnification learning frameworks and visualize feature representations.
• To overcome challenges in understanding cross-magnification information propagation.

Main Methods:

• Developed a deep multi-magnification similarity learning (DSML) approach.
• Utilized a similarity cross-entropy loss function for cross-magnification information similarity.
• Conducted experiments with various network backbones and magnification combinations.
• Investigated interpretability through visualization techniques.

Main Results:

• DSML achieved outstanding classification performance on nasopharyngeal carcinoma and breast cancer datasets.
• The method demonstrated higher area under the curve, accuracy, and F-score compared to existing methods.
• Visualization confirmed the effectiveness and interpretability of the multi-magnification approach.

Conclusions:

• The proposed DSML approach effectively integrates information from multiple magnifications for improved histopathological image classification.
• DSML offers enhanced interpretability, bridging the gap between low-dimension cell-level and high-dimension tissue-level features.
• The study highlights the significant potential of multi-magnification learning in digital pathology and computer-aided diagnosis.