Related Experiment Video
Updated: May 28, 2025
13:01
Industrialized, Artificial Intelligence-guided Laser Microdissection for Microscaled Proteomic Analysis of the Tumor Microenvironment
Published on: June 3, 2022
3.6K
Context-guided segmentation for histopathologic cancer segmentation.
Jeremy Juybari1,2, Josh Hamilton1, Chaofan Chen3
1CompuMAINE Lab, Department of Chemical and Biomedical Engineering, University of Maine, Orono, 04469, USA.
Scientific Reports
|February 13, 2025
Summary
This study introduces a dual-encoder model for cancer diagnosis, mimicking pathologists' zoom technique. The model uses both detailed and contextual views of tissue for improved pixel-wise segmentation accuracy.
Area of Science:
- Digital Pathology
- Computational Biology
- Medical Image Analysis
Background:
- Histological tissue inspection is the gold standard for cancer diagnosis.
- Pathologists analyze tissue samples using varying levels of magnification.
Purpose of the Study:
- To develop a dual-encoder model for cancer diagnosis that mimics pathologist's zoom-in/zoom-out approach.
- To improve pixel-wise segmentation of cancerous tissue by integrating contextual and detailed views.
Main Methods:
- A dual-encoder model was proposed, processing tissue views at different magnifications concurrently.
- The model utilizes two encoder branches for detail and context resolutions.
- Unique weight initialization for cross-attention was introduced to integrate contextual information.
Main Results:
- The dual-encoder model demonstrated improved performance on the Camelyon16 dataset.
- An increase in Area Under the Curve (AUC) from 0.31% to 0.92% was observed.
- Cancer Dice score improved by 4.09% to 6.81% compared to single-view models.
Conclusions:
- Integrating contextual information alongside detailed views significantly enhances cancer segmentation.
- The proposed dual-encoder model offers a promising approach for automated cancer diagnosis in digital pathology.
