Related Experiment Video
Updated: Jul 9, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
1.9K
PATrans: Pixel-Adaptive Transformer for edge segmentation of cervical nuclei on small-scale datasets
Hexuan Hu1, Jianyu Zhang1, Tianjin Yang1
1Key Laboratory of Water Big Data Technology of Ministry of Water Resources, Hohai University, Nanjing, 211100, PR China; College of Computer and Information, Hohai University, Nanjing, 211100, PR China.
Computers in Biology and Medicine
|December 7, 2023
Summary
Pixel Adaptive Transformer (PATrans) enhances cervical nuclei segmentation on small datasets by adaptively tuning pixels. This approach overcomes Transformer limitations with limited data, improving edge segmentation accuracy.
Area of Science:
- Medical image analysis
- Computer vision
- Deep learning
Background:
- Transformers excel in visual tasks but struggle with small medical datasets.
- Limited nuclei pixels hinder Transformer's ability to model local structures and scale variations.
Purpose of the Study:
- To propose Pixel Adaptive Transformer (PATrans) for improved nuclei edge segmentation on small datasets.
- To address information loss and scale variation issues in Transformer models for medical imaging.
Main Methods:
- Developed Consecutive Pixel Patch (CPP) to embed multi-scale context into image patches, ensuring scale invariance.
- Introduced Pixel Adaptive Transformer Block (PATB) for data-dependent pixel relationship modeling across feature maps.
- Enabled adaptive reduction of irrelevant pixel interference by learning local features and global dependencies.
Main Results:
- PATrans demonstrated superior performance in nuclei edge segmentation compared to existing methods.
- The model effectively handled scale variations and improved semantic consistency.
- Experiments on ISBI and Herlev datasets validated the model's effectiveness.
Conclusions:
- PATrans significantly improves nuclei segmentation on small datasets by adaptively tuning pixels.
- The proposed CPP and PATB effectively enhance Transformer's capability in medical image analysis.
- PATrans offers a promising solution for Transformer-based medical image segmentation challenges.
