Radiation dose-aware sinogram knowledge library transformer with feature modulation for low-dose medical image segmentation.
Authors
Affiliations (4)
Affiliations (4)
- Department of Artificial Intelligence Convergence, Chonnam National University, 77 Yongbong-ro, Buk-gu, Gwangju, 61186, Republic of Korea. Electronic address: [email protected].
- Department of Artificial Intelligence Convergence, Chonnam National University, 77 Yongbong-ro, Buk-gu, Gwangju, 61186, Republic of Korea. Electronic address: [email protected].
- Department of Artificial Intelligence Convergence, Chonnam National University, 77 Yongbong-ro, Buk-gu, Gwangju, 61186, Republic of Korea. Electronic address: [email protected].
- Department of Artificial Intelligence Convergence, Chonnam National University, 77 Yongbong-ro, Buk-gu, Gwangju, 61186, Republic of Korea. Electronic address: [email protected].
Abstract
Low-dose computed tomography (LDCT) and low-dose positron emission tomography (LDPET) enable shorter acquisition times and lower radiation exposure. However, dose reduction leads to low-count data, which increases noise and variance and can introduce structured artifacts. These degradations hinder reliable clinical interpretation by reducing lesion contrast and obscuring lesion boundaries, thereby compromising segmentation accuracy. Although deep learning approaches have been proposed to reconstruct and segment low-dose images to full-dose quality, many existing methods address dose variation without explicitly modeling the relationship between dose reduction and image degradation. Moreover, some methods assume externally provided dose metadata as a conditioning input, which can limit applicability and generalization when such metadata are missing, non-standardized, or defined inconsistently across datasets and protocols. Therefore, this work proposes the radiation dose-aware sinogram knowledge library transformer with feature modulation for low-dose medical image segmentation (SinoDose). The main novelty of SinoDose is that it estimates an observation-driven visual relative dose (VR-dose) directly from sinogram-domain internal cues, rather than relying on external dose metadata. In addition, a learnable sinogram knowledge library (SKL) injects acquisition-domain periodic priors to compensate for structures that collapse or become distorted under extremely low-dose conditions. The estimated VR-dose and its context vector are then used in dose-guided calibrated FiLM to enable dose-adaptive joint reconstruction and lesion segmentation in a unified framework. Experiments on the annotated AutoPET and KiTS datasets show that SinoDose consistently improves DSC, HD95, PSNR, and rPSNR over competing baselines. Additional evaluation on the real low-dose UDPET and the real-world LDCT datasets further supports reconstruction robustness under practical acquisition conditions. The code is available at https://github.com/yeongjong0220/DoseSino.