Related Experiment Video
Updated: Sep 6, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Adaptive Cross Entropy for ultrasmall object detection in Computed Tomography with noisy labels
Hedan Chen1, Weimin Tan1, Jichun Li1
1School of Computer Science, Shanghai Key Laboratory of Intelligent Information Processing, Fudan University, Shanghai, 200433, China.
Computers in Biology and Medicine
|July 1, 2022
Summary
Researchers developed a new method for detecting ultrasmall objects, even with imperfect data. This approach improves diagnostic accuracy for conditions like stapedial otosclerosis using a novel Adaptive Cross Entropy (ACE) loss function and W-Net architecture.
Area of Science:
- Medical Imaging
- Computer Vision
- Biomedical Engineering
Background:
- Ultrasmall object detection is challenging due to limited datasets.
- Stapedial otosclerosis detection requires specialized imaging analysis.
- Noisy labels are a common issue in real-world datasets.
Purpose of the Study:
- To address the scarcity of ultrasmall object detection datasets.
- To improve stapedial otosclerosis detection accuracy.
- To develop a robust method for learning with noisy labels.
Main Methods:
- Collected the largest stapedial otosclerosis dataset.
- Proposed Adaptive Cross Entropy (ACE) loss for noisy labels.
- Introduced W-Net auxiliary backbone for high-resolution features.
Main Results:
- ACE loss significantly boosts diagnostic performance with noisy labels.
- W-Net enhances extraction of high-resolution representations for ultrasmall objects.
- The combined approach achieves superior results in ultrasmall object detection.
Conclusions:
- The proposed ACE loss and W-Net are effective for ultrasmall object detection.
- This work advances research in learning with noisy labels.
- The findings have implications for medical image analysis and diagnosis.
