Generalized deep learning for histopathology image classification using supervised contrastive learning.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Digital pathology

Background:

• Cancer remains a leading global cause of death, underscoring the need for precise diagnostic tools.
• Histopathological image analysis is vital for cancer diagnosis but faces challenges from human error and variability.
• HistopathAI is introduced as a hybrid network to enhance diagnostic precision and efficiency in clinical pathology.

Purpose of the Study:

• To demonstrate HistopathAI's capability to improve histopathological image classification accuracy.
• To validate the effectiveness of supervised contrastive learning (SCL) and hybrid deep feature fusion (HDFF) in enhancing diagnostic precision.
• To show improved performance on imbalanced datasets.

Main Methods:

• HistopathAI integrates EfficientNetB3 and ResNet50 features via HDFF for comprehensive image representation.
• A sequential approach transitions from feature learning to classifier learning, inspired by contrastive learning principles.
• The model combines SCL for feature representation with cross-entropy loss for classification.

Main Results:

• HistopathAI achieved state-of-the-art classification accuracy on multiple public and private histopathology datasets.
• Superior performance was observed in both binary and multiclass classification tasks.
• Statistical analysis confirmed significant improvements over baseline models.

Conclusions:

• HistopathAI provides a robust solution for histopathology image classification, boosting diagnostic accuracy.
• The framework supports the adoption of digital pathology and has the potential to improve patient outcomes.
• The developed code is publicly available to facilitate broader clinical application and research.