Related Experiment Video
Updated: Jun 19, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Deblurring masked image modeling for ultrasound image analysis
Qingbo Kang1, Qicheng Lao2, Jun Gao3
1Department of Ultrasonography, West China Hospital, Sichuan University, Chengdu, Sichuan, 610041, China; West China Biomedical Big Data Center, West China Hospital, Sichuan University, Chengdu, Sichuan, 610041, China; Shanghai Artificial Intelligence Laboratory, Shanghai, 200030, China.
This study introduces a specialized masked image modeling (MIM) approach for ultrasound images, incorporating deblurring to enhance detail recovery. The novel method achieves state-of-the-art results in various diagnostic tasks, improving ultrasound image analysis.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- Large pretrained vision foundation models, particularly those using masked image modeling (MIM), show promise in various tasks.
- MIM applications in ultrasound imaging are underexplored, and existing methods do not address ultrasound's unique characteristics, like high noise-to-signal ratio.
Purpose of the Study:
- To develop a specialized masked image modeling (MIM) approach tailored for ultrasound imaging.
- To address the limitations of current MIM methods in handling ultrasound's high noise-to-signal ratio and image characteristics.
Main Methods:
- Proposed a deblurring masked image modeling (MIM) approach incorporating a deblurring task into the pretraining proxy task.
- Employed a multi-scale hierarchical encoder to capture both local and global contextual information.
- Conducted pretraining on 280,000 ultrasound images.
Main Results:
- The deblurring MIM approach effectively recovers subtle details crucial for downstream analysis.
- The multi-scale encoder improved performance, particularly on pixel-wise tasks like segmentation.
- Achieved state-of-the-art performance on various disease diagnoses (nodule, Hashimoto's thyroiditis) and task types (classification, segmentation).
Conclusions:
- The proposed deblurring MIM is effective for ultrasound image analysis.
- This work presents a specialized vision foundation model for ultrasound, demonstrating the potential of MIM in this domain.
