Related Experiment Video
Updated: Jun 17, 2025
07:32
Author Spotlight: Investigating Immune Cell Dynamics in the Tumor Microenvironment — Challenges and Innovations in Cancer Prognosis
Published on: April 12, 2024
1.3K
Robust ROI Detection in Whole Slide Images Guided by Pathologists' Viewing Patterns
Fatemeh Ghezloo1, Oliver H Chang2, Stevan R Knezevich3
1Paul G. Allen School of Computer Science and Engineering, University of Washington, Seattle, WA, USA. fghezloo@uw.edu.
Journal of Imaging Informatics in Medicine
|August 9, 2024
Summary
This study introduces a novel deep learning system that uses pathologist viewing heatmaps to improve diagnostic accuracy in whole slide images. The approach enhances region detection for computer-aided diagnosis without requiring manual annotations.
Area of Science:
- Digital pathology
- Artificial intelligence in medicine
- Computer-aided diagnosis
Background:
- Pathologist expertise is crucial for accurate diagnosis from whole slide images.
- Annotation of critical regions in pathology images is time-consuming and requires specialized knowledge.
- Deep learning models benefit from understanding pathologist viewing patterns to focus on diagnostically relevant areas.
Purpose of the Study:
- To develop a deep learning system that integrates pathologist viewing patterns (heatmaps) to guide region of interest detection.
- To improve the performance of computer-aided diagnosis systems by leveraging domain expertise.
- To reduce the need for manual annotations in training deep learning models for pathology.
Main Methods:
- Generating heatmaps from pathologist viewing data to represent diagnostic focus areas.
- Training a U-Net deep learning model with a ResNet-18 encoder, guided by these heatmaps.
- Evaluating the model on a skin biopsy dataset for melanoma diagnosis and comparing it to traditional methods.
- Conducting a clinical evaluation with dermatopathologists to assess the model's performance and relevance.
Main Results:
- The proposed system demonstrated superior performance over traditional methods, with significant increases in precision (20%), recall (11%), F1-score (22%), and Intersection over Union (12%).
- The U-Net model effectively identified regions of interest, mimicking pathologist diagnostic behavior.
- Clinical evaluation confirmed the model's ability to replicate expert viewing patterns and identify critical diagnostic regions.
Conclusions:
- Incorporating heatmaps of pathologist viewing patterns as supplementary signals enhances deep learning-based computer-aided diagnosis systems.
- This approach shows promise for improving diagnostic accuracy and efficiency in digital pathology.
- The method streamlines annotation processes and can aid in training new pathologists by highlighting key diagnostic areas.
