Metric Learning in Histopathological Image Classification: Opening the Black Box

Area of Science:

• Digital Pathology
• Machine Learning in Medicine
• Computational Biology

Background:

• Histopathology image analysis is crucial for disease diagnosis.
• Manual classification is time-consuming and prone to fatigue.
• Machine learning offers automated solutions to assist pathologists.

Purpose of the Study:

• To apply metric learning, specifically triplet networks, for histopathology image classification.
• To develop a system that provides interpretable information beyond simple classification.
• To improve the decision-making process for physicians analyzing medical images.

Main Methods:

• Utilized triplet networks to learn image representations in an embedding space.
• Grouped similar images and separated dissimilar ones based on learned distances.
• Evaluated the model on the BreakHis dataset, containing breast cancer histopathology images.

Main Results:

• Achieved effective class separation in the embedding space.
• Demonstrated strong classification performance at the patient level.
• Provided interpretable results based on image similarity and dissimilarity metrics.

Conclusions:

• Metric learning, particularly with triplet networks, is a valuable tool for histopathology image classification.
• The proposed method offers interpretable insights, differentiating it from recent approaches.
• This technique can significantly aid physicians in the diagnosis of breast cancer.