Center-Focused Affinity Loss for Class Imbalance Histology Image Classification.

Area of Science:

• Computational pathology
• Deep learning in medical imaging
• Histopathology image analysis

Background:

• Manual analysis of Whole Slide Images (WSIs) for cancer diagnosis is time-consuming.
• Deep learning in computational pathology offers potential for efficient tumor-microenvironment analysis.
• Existing deep learning models struggle with fine-grained histopathology datasets due to imbalanced data and conventional loss functions.

Purpose of the Study:

• To develop a novel loss function for deep learning models to better handle imbalanced histopathology datasets.
• To improve the distinctiveness of representational embeddings for similarly textured WSIs.
• To enhance the accuracy of early-stage cancer diagnosis through improved computational pathology methods.

Main Methods:

• Proposed a novel center-focused affinity loss (CFAL) function.
• CFAL constructs uniform class prototypes, penalizes difficult samples, minimizes intra-class variations, and emphasizes minority class features.
• Evaluated CFAL on imbalanced breast and colon cancer datasets.

Main Results:

• The proposed CFAL function demonstrated superior discrimination abilities compared to ArcFace, CosFace, and Focal loss.
• CFAL outperformed several state-of-the-art (SOTA) methods in histology image classification on both datasets.
• The loss function effectively addresses challenges posed by imbalanced data distributions in histopathology.

Conclusions:

• The novel CFAL loss function significantly improves deep learning model performance for histology image classification.
• CFAL offers a promising solution for accurate and efficient cancer diagnosis using computational pathology.
• This approach has the potential to aid pathologists in early cancer detection and improve patient survival rates.