Related Experiment Video
Updated: Jun 6, 2026
08:30
X-ray Dose Reduction through Adaptive Exposure in Fluoroscopic Imaging
Published on: September 11, 2011
Radiation dose-aware sinogram knowledge library transformer with feature modulation for low-dose medical image
Yeong Jong Lee1, Do Hyun Ki1, Seung-Hyeok Back1
1Department of Artificial Intelligence Convergence, Chonnam National University, 77 Yongbong-ro, Buk-gu, Gwangju, 61186, Republic of Korea.
Computers in Biology and Medicine
|June 4, 2026
Summary
This study introduces SinoDose, a novel deep learning method for low-dose medical image segmentation. SinoDose accurately segments lesions by estimating dose from image data, improving accuracy without external dose information.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Low-dose computed tomography (LDCT) and low-dose positron emission tomography (LDPET) reduce radiation exposure but introduce noise and artifacts, degrading image quality.
- Image degradations in low-dose scans compromise lesion contrast, boundary definition, and segmentation accuracy, hindering clinical interpretation.
- Existing deep learning methods often fail to explicitly model dose-reduction effects or rely on external, potentially inconsistent, dose metadata.
Purpose of the Study:
- To develop a novel deep learning framework, SinoDose, for accurate low-dose medical image segmentation.
- To enable dose-adaptive joint reconstruction and segmentation without requiring external dose metadata.
- To improve lesion segmentation accuracy and image quality in low-dose scans.
Main Methods:
- Proposed SinoDose, a transformer-based method utilizing a sinogram knowledge library (SKL) and feature modulation.
- Introduced an observation-driven visual relative dose (VR-dose) estimation directly from sinogram-domain cues.
- Employed dose-guided calibrated FiLM (feature-wise linear modulation) for joint reconstruction and segmentation.
Main Results:
- SinoDose demonstrated consistent improvements in Dice Similarity Coefficient (DSC), Hausdorff Distance 95% (HD95), Peak Signal-to-Noise Ratio (PSNR), and regional PSNR (rPSNR) on AutoPET and KiTS datasets.
- The method showed robust performance on real low-dose UDPET and real-world LDCT datasets, validating its practical applicability.
- SinoDose outperformed competing baseline methods in low-dose medical image segmentation tasks.
Conclusions:
- SinoDose effectively addresses image degradation in low-dose scans by learning dose information internally.
- The proposed method enhances lesion segmentation accuracy and image reconstruction quality without external dose metadata.
- SinoDose offers a robust and generalizable solution for low-dose medical image analysis.
