An Adaptive Attention-Driven Quadruplet Deep Hashing Method for Retrieving Histopathological Images

Area of Science:

• Digital Pathology
• Computer Vision
• Machine Learning

Background:

• Histopathological image retrieval aids disease diagnosis and treatment planning.
• High-dimensional features in these images present complexity and inefficiency challenges.
• Deep hashing methods offer solutions but face the vanishing gradient problem.

Purpose of the Study:

• To develop an adaptive quadruplet deep hashing model for enhanced histopathological image retrieval.
• To address the vanishing gradient issue in deep hashing models.
• To improve the efficiency and accuracy of histopathological image retrieval systems.

Main Methods:

• Utilized four deep hashing models with identical structures to generate hash codes.
• Implemented a novel adaptive quadruplet loss function for training hash codes.
• Incorporated an attention module within a convolutional neural network architecture to boost feature extraction.

Main Results:

• The adaptive structure demonstrated improved retrieval performance.
• A novel hash layer effectively mitigated the vanishing gradient issue.
• The attention module enhanced feature extraction efficiency.
• Achieved highest mean average precision (MAP) on Kather, Kimia Path960, and Kimia Path24C datasets (approx. 0.9940, 0.9983, 0.9968).

Conclusions:

• The proposed adaptive quadruplet deep hashing model significantly outperforms existing hashing techniques for histopathological image retrieval.
• The model successfully addresses key challenges including high dimensionality and vanishing gradients.
• The integration of an attention mechanism further refines feature extraction, leading to superior performance.