Learning regions of interest from low level maps in virtual microscopy

Area of Science:

• Digital Pathology
• Computational Pathology
• Medical Image Analysis

Background:

• Virtual microscopy (VS) enhances pathology lab workflows but is hindered by large file sizes.
• Identifying Regions of Interest (RoIs) is crucial for VS navigation and data compression.

Purpose of the Study:

• To present a novel method for automatically determining RoIs in virtual slides.
• To improve VS navigation and compression efficiency by leveraging pathologist expertise.

Main Methods:

• Virtual slides are divided into blocks, each represented by low-level features (color, intensity, orientation, texture).
• Pathologists select example blocks as relevant (positive) or irrelevant (negative) instances.
• Similarity and dissimilarity maps are generated from these examples.
• A normalization process integrates maps, prioritizing regions with high similarity maxima.
• Each image region receives a relevance score based on proximity to positive and negative examples.

Main Results:

• The method was evaluated on 8 virtual slides from diverse tissues.
• Pathologists navigated the VS using the generated relevance scores.
• Precision-recall measurements showed an average precision of 55% and a recall of approximately 38% during navigation.

Conclusions:

• The proposed method effectively identifies relevant regions of interest in virtual slides.
• This approach has the potential to significantly improve the efficiency of virtual microscopy navigation and data handling in pathology.
• The integration of pathologist-selected examples provides a powerful tool for automated RoI determination.