Related Experiment Video
Updated: Jun 24, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
Adaptive distribution-aware transformer for multi-scale visual representation learning on imbalanced and
Sakib Ahammed1, Xia Cui1, Wenqi Lu1
1Department of Computing and Mathematics, Manchester Metropolitan University, The Dalton Building, Chester Street, Manchester, M1 5GD, UK.
Medical Image Analysis
|June 22, 2026
Summary
AdaptiveViT, a new deep learning model, effectively handles imbalanced medical datasets and low-resolution images. This hybrid CNN-Transformer architecture improves classification accuracy for rare conditions like melanoma.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Deep learning models face challenges with class imbalance and low-resolution medical images.
- Minority-class features and critical spatial details are often underrepresented, impacting diagnostic accuracy.
Purpose of the Study:
- To introduce Adaptive Distribution-aware Vision Transformer (AdaptiveViT), a novel hybrid CNN-Transformer architecture.
- To address class imbalance and image resolution variability in medical image classification.
Main Methods:
- AdaptiveViT unifies fine-grained local feature extraction (CNN) with global contextual modeling (Transformer).
- It incorporates a distribution-aware modulation mechanism and a Distribution-aware Adaptive (DA) Loss to enhance minority-class sensitivity.
- Experiments were conducted on skin lesion and gastrointestinal endoscopy datasets with varying resolutions and imbalance ratios.
Main Results:
- AdaptiveViT outperformed state-of-the-art baselines in F1 and AUC scores across multiple datasets.
- The model demonstrated stable convergence across different levels of class imbalance.
- Validation on endoscopy data confirmed AdaptiveViT's domain-agnostic generalization capabilities.
Conclusions:
- AdaptiveViT establishes a robust hybrid framework for medical image classification, particularly under class imbalance and resolution variability.
- The approach offers improved diagnostic reliability for underrepresented conditions in medical imaging.
