Optimization of Image Quality and Radiation Dose in Digital Radiography using Deep Learning-Based Image Reconstruction

Project Abstract

This study investigates the simultaneous enhancement of diagnostic image quality and reduction of patient radiation dose in digital radiography (DR) through the integration of deep learning-based image reconstruction techniques with established imaging physics models. A multi-stage framework was developed to address the core challenges of DR noise amplification at low-dose exposure, limited spatial resolution, and artifact susceptibility due to patient motion and hardware imperfections. The proposed approach combines physics-informed neural networks (PINNs) with model-based iterative reconstruction (MBIR) to enforce data fidelity while leveraging learned priors for edge preservation and texture fidelity. A comprehensive dataset comprising clinical DR images acquired at varied dose levels, along with corresponding high-dose references and phantom measurements, was assembled to train and validate the system. The deep learning component employs a cascaded residual architecture with attention mechanisms to adaptively denoise and super-resolve projection images, while a differentiable forward model enables end-to-end optimization of dose-qualit y trade-offs. Regularization terms derived from the Beer-Lambert law and detector response are embedded within the training objective to maintain physically plausible reconstructions. To ensure generalizability, domain adaptation strategies were implemented to reconcile differences across manufacturers, imaging protocols, and anatomical regions. Quantitative evaluation utilized standardized metrics including peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and contrast-to-noise ratio (CNR), alongside clinically oriented assessments by radiologists focusing on lesion detectability and artifact suppression. The study also introduces a robust dose reduction framework, quantifying potential reductions in Entrance Surface Dose (ESD) and effective dose while preserving or enhancing diagnostic accuracy. Reproducibility is addressed through open-sourced code, a modular data processing pipeline, and a reproducible evaluation protocol with stratified cross-validation across different anatomical landmarks. Results demonstrate that the proposed deep learning–assisted reconstruction yields substantial improvements in image quality at reduced dose levels, achieving comparable or superior SSIM and PSNR relative to conventional MBIR methods at standard doses. Notably, edge sharpness and texture fidelity in high-frequency regions are preserved, with a marked reduction in motion-related artifacts. Clinical validation indicates improved conspicuity of osseous and soft-tissue structures, enhanced lesion delineation, and reduced need for repeat radiographs due to degraded image quality. Additionally, the framework exhibits resilience to variations in exposure parameters and patient size, suggesting its applicability across diverse clinical workflows. The study contributes a rigorous methodology for balancing dose minimization with diagnostic assurance, a validated deep learning–augmented reconstruction paradigm, and practical guidelines for integration into radiography departments. Future work will explore real-time deployment on edge devices, multi-view fusion for 3D tomosynthesis, and the incorporation of active learning to continually refine model performance with accruing clinical data.

Project Overview

What This Project Is About

---

A straightforward exploration of how deep learning can improve X-ray images. The project looks at methods that reconstruct clearer pictures from digital radiography data while reducing the radiation dose patients must receive.

---

---

The Problem It Addresses

---

Low-dose X-rays tend to be noisy and less sharp, which can obscure important details. Higher doses improve image quality but raise patient risk. The project investigates a way to keep dose low while maintaining or improving image clarity using machine learning.

---

---

Objectives of the Project

---

1. Understand the basics of digital radiography and image quality metrics.
2. Learn how deep learning can reconstruct medical images from noisy data.
3. Develop a simple model to enhance low-dose radiographs.
4. Evaluate image quality with objective metrics and, if possible, expert review.

---

---

What You Will Do Step by Step

---

1) Review background materials on radiography and image reconstruction. 2) Collect or simulate paired low-dose and standard-dose images. 3) Build a basic neural network model for image enhancement. 4) Train and validate the model using appropriate metrics. 5) Compare results to traditional denoising methods. 6) Document findings and discuss practical implications.

---

---

---

---

Expected Outcome

The project should produce a working demonstration that showing clearer images from lower-dose scans, along with a discussion of trade-offs, limitations, and potential benefits for patient safety and workflow in clinics.