Related Experiment Video
Updated: Jul 1, 2025
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
42.9K
[Segmentation of Mass in Mammogram Using Gaze Search Patterns].
Eiichiro Okumura1, Hideki Kato2, Tsuyoshi Honmoto3
1Department of Radiological Technology, Faculty of Health Sciences, Tsukuba International University.
Nihon Hoshasen Gijutsu Gakkai Zasshi
|March 13, 2024
Summary
This study explored using eye-tracking focus images to improve mammogram mass segmentation. Combining UNIT-generated focus images with original mammograms achieved segmentation performance comparable to using original mammograms alone.
Area of Science:
- Medical Imaging
- Computer Vision
- Radiology
Context:
- Mammography interpretation is challenging for detecting small or obscured lesions in dense breast tissue.
- Computer-aided detection (CAD) systems are increasingly used to aid radiologists.
- Accurate segmentation of masses is crucial for effective CAD systems.
Purpose:
- To accurately segment mammographic masses using focus images generated by an eye-tracking device.
- To evaluate the performance of different deep learning models (auto-encoder, Pix2Pix, UNIT) in generating focus images.
- To assess the combined efficacy of focus images and original mammograms for mass segmentation.
Summary:
- Focus images were generated using auto-encoder, Pix2Pix, and UNIT models trained on radiologist eye-tracking data.
- The UNIT model achieved a higher Dice coefficient (0.64±0.14) compared to auto-encoder and Pix2Pix.
- Combining UNIT-generated focus images with original mammograms yielded a Dice coefficient of 0.66±0.15, comparable to using original mammograms alone.
Impact:
- This research explores novel methods for enhancing mammogram analysis through eye-tracking data integration.
- The findings suggest potential for improving the accuracy of computer-aided detection systems.
- Further research with larger datasets is needed to optimize segmentation performance.
