Related Experiment Video
Updated: Sep 25, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
659
Endoscopy Artefact Detection by Deep Transfer Learning of Baseline Models
Tang-Kai Yin1, Kai-Lun Huang2, Si-Rong Chiu2
1Department of Computer Science and Information Engineering, National University of Kaohsiung, No. 700, Kaohsiung University Rd., Nan-Tzu Dist., 811, Kaohsiung, Taiwan. tkyin@nuk.edu.tw.
Journal of Digital Imaging
|April 28, 2022
Summary
Deep learning models, including EfficientDet-D2, effectively detect eight types of artefacts in endoscopic videos. This research improves cancer diagnosis by enhancing image quality in medical imaging.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- Endoscopic video frames often contain artefacts that hinder accurate cancer diagnosis.
- Artefacts such as specularity, bubbles, saturation, contrast, blood, instruments, blur, and imaging issues complicate medical interpretation.
Purpose of the Study:
- To apply deep learning for detecting eight common artefacts in endoscopic video frames.
- To evaluate the performance of state-of-the-art deep learning models in artefact detection.
Main Methods:
- Utilized transfer learning with pre-trained parameters and fine-tuning.
- Applied two leading deep learning methods: Faster Region-based Convolutional Neural Networks (Faster R-CNN) and EfficientDet.
- Trained and tested models on the Endoscopy Artefact Detection and Segmentation (EAD2020) dataset, using phase I for training and phase II for testing.
Main Results:
- EfficientDet-D2 achieved a competitive score of 0.2008 (mAPd[Formula: see text]0.6+mIoUd[Formula: see text]0.4), outperforming Faster R-CNN, YOLOv3, and RetinaNet.
- The EfficientDet-D2 model demonstrated state-of-the-art performance without advanced techniques like test-time augmentation.
- Results were competitive with the top leaderboard scores, despite using a different testing subset.
Conclusions:
- Deep learning, particularly EfficientDet-D2, offers a robust solution for detecting artefacts in endoscopy.
- The proposed combination of EfficientDet-D2 with data augmentation and pre-trained parameters enhances artefact detection capabilities.
- This approach has the potential to improve the accuracy and reliability of cancer diagnosis through clearer endoscopic imaging.
