MuCoSA: Multi-contextual similarity assessment for histopathology image search

Area of Science:

• Digital pathology
• Computational pathology
• Histopathology image analysis

Background:

• Histological diagnosis relies on pattern recognition, but current digital pathology tools have limitations in capturing multi-scale tissue morphology.
• Existing image retrieval frameworks often use single-magnification patches, hindering comprehensive analysis.

Purpose of the Study:

• To introduce the Progressive Regional Image Sequence by Magnification (PRISM) and the Multi-Contextual Similarity Assessment (MuCoSA) framework for enhanced histopathology image retrieval.
• To evaluate MuCoSA's effectiveness in capturing multi-magnification contextual information without additional fine-tuning of pre-trained feature encoders.

Main Methods:

• Developed PRISM to represent tissue morphology across multiple magnifications.
• Implemented MuCoSA, a PRISM-based retrieval framework utilizing pre-trained encoders.
• Constructed reference and query datasets using lung adenocarcinoma histology from Samsung Medical Center and The Cancer Genome Atlas.
• Calculated similarity by averaging cosine similarities of corresponding patches across all magnifications.

Main Results:

• MuCoSA with multi-magnification significantly outperformed single-magnification methods, achieving high F1-scores, mAP@5, and mMV@5.
• For example, MuCoSA with eight magnifications achieved an F1-score of 0.8051, significantly higher than the baseline (p < 0.0001).
• Search results aligned with pathologists' morphological interpretations, confirmed by confusion matrices and visual inspection.

Conclusions:

• MuCoSA provides a simple, efficient, and effective framework for improving histopathology image search using multi-magnification PRISM without fine-tuning.
• This approach can significantly enhance diagnostic accuracy and consistency for pathologists, reducing observer variability.