A Practical Weakly Supervised Framework for Dose-Up Translation of Low-Enhanced CT Under Clinical Acquisition...

Area of Science:

• Medical Imaging
• Computational Imaging
• Veterinary Radiology

Background:

• Low-dose contrast-enhanced computed tomography (CT) aims to reduce toxicity but suffers from image quality issues like uneven attenuation and spatial misalignment.
• These limitations hinder effective dose-up translation, especially in veterinary abdominal CT due to motion and variability.
• Existing methods struggle with spatial misalignment and limited paired low-dose data.

Purpose of the Study:

• To develop a computational framework for dose-up translation of low-dose contrast CT images to full-dose quality.
• To address challenges of spatial misalignment and limited paired data in veterinary CT.
• To improve diagnostic quality and anatomical fidelity in enhanced CT images.

Main Methods:

• A weakly aligned enhancement framework using registration-based pseudo-references (deformable alignment + feature correspondence).
• Structure-preserving translation with multi-scale consistency and edge-aware regularization.
• A two-stage knowledge transfer strategy using pre-contrast data to overcome limited low-dose datasets.

Main Results:

• Achieved region-level contrast-to-noise ratio improvements up to 31.5% (e.g., CVC: 5.55 to 8.38, p < 0.05).
• Demonstrated improved structural fidelity, distributional realism, and vascular conspicuity compared to baseline methods.
• Outperformed paired, unpaired, and synthetic-pairing baselines in quantitative and qualitative evaluations.

Conclusions:

• Dose-up translation of low-enhanced CT is effectively framed as a weakly aligned domain adaptation problem.
• The proposed method enables practical image translation under realistic clinical acquisition variability.
• The framework enhances diagnostic utility of low-dose CT by approximating full-dose quality.