Related Experiment Video
Updated: Oct 4, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
On Smart Gaze Based Annotation of Histopathology Images for Training of Deep Convolutional Neural Networks
IEEE Journal of Biomedical and Health Informatics
|February 7, 2022
Summary
Eye gaze annotation significantly speeds up pathology slide labeling for deep learning. This method reduces pathologist time by up to 85% while maintaining good performance for training object detectors.
Area of Science:
- Digital Pathology
- Machine Learning in Medicine
- Histopathology Image Analysis
Background:
- Deep learning for histopathology requires large annotated datasets, which are time-consuming to create.
- Whole slide imaging accelerates data acquisition but manual slide annotation remains a bottleneck.
- Eye gaze tracking offers a potential solution to expedite the slide labeling process.
Purpose of the Study:
- To evaluate the viability and efficiency of eye gaze annotations for training object detection models in histopathology.
- To compare the time investment and performance of gaze-based labeling versus conventional manual labeling methods.
- To discuss challenges and refinement techniques for gaze-based annotations in deep learning.
Main Methods:
- Comparison of time taken for eye gaze labeling versus manual bounding-box and freehand labeling.
- Training deep object detectors using both gaze-labeled and manually labeled datasets.
- Evaluation of model performance on a specific task: localizing Keratin Pearls in oral squamous cell carcinoma.
- Exploration of methods to refine coarse gaze data for improved annotation quality.
Main Results:
- Gaze-based labeling significantly reduces the time required per annotation compared to manual methods.
- Gaze-labeling required 57.6% less time than bounding-box labeling and 85% less time than freehand labeling.
- Deep object detectors trained with gaze-labeled data achieve good performance, comparable to those trained with manual labels.
- Identified challenges in gaze-based annotation and proposed refinement strategies.
Conclusions:
- Eye gaze annotation is a viable and time-efficient method for creating training data in digital pathology.
- This approach can alleviate the bottleneck of manual annotation, enabling larger datasets for deep learning.
- Gaze-based labeling holds promise for accelerating AI development in histopathology applications.
