Self-supervised Pyramid Mask Transformer for Universal Chest CT Prognosis Pre-training
Principal Investigator
Name
Zhexin Chen
Degrees
Ph.D.
Institution
Southeast University
Position Title
Ph.D. Researcher
Email
About this CDAS Project
Study
NLST
(Learn more about this study)
Project ID
NLST-1238
Initial CDAS Request Approval
Apr 29, 2024
Title
Self-supervised Pyramid Mask Transformer for Universal Chest CT Prognosis Pre-training
Summary
The prognosis prediction of chest disease through CT imaging is crucial for diagnosis and treatment. Usually, prognosis prediction involves statistical analysis that requires a large amount of survival data, while the high cost of follow-up results in a limited survival data scale. In this paper, we first propose a Universal Chest CT Transformer (UCTrans) to extract features of varying granularity from chest CT images and facilitate prognostic predictions for various chest diseases through a self-supervised pre-training framework. Initially, in UCTrans, our proposed Feature Wander Module (FWM) is combined with overlapping patch embedding to enable the model to perceive the complex fine structure of chest CT. Subsequently, we introduce pyramid mask modeling, a multi-level masking method, enabling the model to learn multi-level features representation of chest CT. Finally, leveraging prior knowledge from pre-training allows adaptation to various downstream prognosis tasks.
Aims
Experiments demonstrate that our UCTrans outperforms competing methods and exhibits superior generalization across different prognosis tasks. We need to pretrain our model on a larger scale dataset to validate our approach, and we also require CSV files containing the data dictionary to be used as prompts for fine-tuning our model in downstream tasks. This contributes to improving prognostic performance and lays the foundation for basic model research in deep learning for the detection, diagnosis, and prognosis of potential cardiovascular and pulmonary diseases.
Collaborators
Guanyu Yang, Southeast University
Yaolei Qi, Southeast University
Shaotong Song, Southeast University
Ying Zheng, Southeast University
Zijin Zhu, Southeast University