Related Experiment Video
Updated: Jul 16, 2025
14:27
Identification of Disease-related Spatial Covariance Patterns using Neuroimaging Data
Published on: June 26, 2013
15.7K
SMiT: symmetric mask transformer for disease severity detection
Chengsheng Zhang1, Cheng Chen2, Chen Chen3
1The College of Software, Xinjiang University, Urumqi, 830046, China.
Journal of Cancer Research and Clinical Oncology
|September 12, 2023
Summary
A new Symmetric Mask Pre-Training vision Transformer (SMiT) model enhances pathological image analysis for disease diagnosis. This deep learning approach improves accuracy in grading cancer images, aiding physicians in clinical decision-making.
Area of Science:
- Medical Imaging
- Deep Learning
- Computational Pathology
Background:
- Deep learning aids intelligent disease diagnosis, but challenges exist in classifying small or uneven lesion areas in pathological images.
- Traditional Convolutional Neural Network (CNN) models may struggle with feature extraction in complex lesion regions, impacting diagnostic accuracy.
Purpose of the Study:
- To propose a pure transformer framework for accurate diagnostic grading of pathological images.
- To develop a novel Symmetric Mask Pre-Training vision Transformer (SMiT) model to overcome limitations in existing deep learning approaches for medical image analysis.
Main Methods:
- The SMiT model employs symmetric mask pre-training on visual transformers, focusing on high-probability sparsification of image token sequences.
- Pre-training weights are loaded to fine-tune the baseline model, enabling enhanced extraction of detailed features within lesion regions and reducing feature dependency.
Main Results:
- SMiT achieved 92.8% classification accuracy on colorectal cancer histopathological images.
- Validation on the APTOS2019 diabetic retinopathy dataset showed quadratic Cohen Kappa, accuracy, and F1-score of 91.9%, 86.91%, and 72.85%, outperforming CNN-based models by 1-2%.
Conclusions:
- The SMiT model offers a simpler yet effective strategy for improving diagnostic accuracy in pathological image grading.
- This transformer-based approach provides valuable insights for leveraging visual transformers in medical imaging applications and assisting clinical decisions.
