Automated Deep Learning Framework for Low-Dose CT Image Denoising and Dose Reduction Validation in Radiography Diagnostics

Project Abstract

In response to the growing demand for diagnostic accuracy while minimizing radiation exposure, this study presents an automated deep learning framework designed for denoising low-dose computed tomography (LDCT) images and validating dose reduction in radiography diagnostics. The framework integrates a multi-stage neural architecture that combines physics-informed priors with data-driven learning to enhance image quality without compromising clinically relevant features. Specifically, we develop a denoising module based on a residual UNet augmented with attention mechanisms and a physics-consistent loss that preserves edge information and texture fidelity under varying noise distributions associated with reduced-dose acquisitions. A second module focuses on dose reduction validation by correlating reconstructed LDCT images with standard-dose CT references through objective metrics and radiologist-annotated perceptual scores, enabling reliable assessment of diagnostic equivalence and potential dose savings. The dataset comprises multi-center LDCT and standard-dose CT pairs, including diverse anatomical regions and scanner models to ensure generalizability. Preprocessing steps address intensity normalization, patient motion correction, and slice-wise denoising compatibility, while data augmentation strategies mitigate limited LDCT samples. The training regime employs a two-tier optimization (i) a reconstruction-centric loss that emphasizes structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), and perceptual quality via a deep feature loss derived from a pre-trained network; and (ii) a physics-consistency loss that enforces adherence to the forward CT model, including system blur and noise statistics, to avoid introducing artifacts that could degrade clinical interpretability. An unsupervised/weakly supervised extension leverages unpaired LDCT data through cycle-consistent generative adversarial constraints to broaden applicability. Validation proceeds through both quantitative and qualitative analyses. Quantitative metrics include SSIM, PSNR, normalized root mean square error (NRMSE), structural similarity of detected lesions, and radiomics feature stability under denoising. Qualitative evaluation involves blinded radiologist assessments focusing on lesion conspicuity, edge sharpness, noise suppression, and overall diagnostic confidence. The study includes a dose reduction scenario analysis, where varying milliampere-seconds (mAs) settings are simulated and the framework’s ability to recover diagnostically meaningful information is rigorously tested. Computational efficiency is addressed by implementing model pruning and quantization to achieve near real-time inference on clinical workstations without sacrificing performance. Results demonstrate that the proposed framework substantially improves LDCT image quality relative to conventional denoising approaches, enabling equivalent diagnostic interpretations at reduced radiation doses. Improvement is observed in lesion visibility, reduced artifacts, and preservation of quantitative radiomics signatures essential for tumor characterization and treatment planning. The dose reduction validation indicates a potential for substantial patient dose savings with maintained diagnostic integrity across multiple scanner models and clinical scenarios. The work also provides a reproducible pipeline for integrating deep learning-based denoising with radiology workflows, including a validation protocol and open-access codebase to facilitate broader adoption and further research in dose-optimized radiographic diagnostics.

Project Overview

What This Project Is About

---

A simple exploration of how computer algorithms can improve CT images taken at lower radiation doses. The project looks at methods that clean up noise and blur in images so doctors can still diagnose accurately while exposing patients to less radiation.

---

---

The Problem It Addresses

---

Lowering radiation in CT scans often makes images noisy and harder to read. This can reduce diagnostic confidence and may require repeat scans. The project seeks ways to restore image quality without increasing dose.

---

---

Objectives of the Project

---

1. Understand the balance between image quality and radiation dose.
2. Learn a basic deep learning approach to denoise CT images.
3. Evaluate how well the method preserves important clinical details.
4. Test different image quality metrics to assess results.
5. Suggest practical guidelines for clinical use.

---

---

What You Will Do Step by Step

---

Step 1: Study CT imaging basics and dose concepts. Step 2: Collect or simulate low-dose and standard-dose CT image pairs. Step 3: Implement a simple denoising model. Step 4: Train and validate the model on the data. Step 5: Compare image quality using objective metrics. Step 6: Analyze how changes in dose affect results. Step 7: Discuss limitations and clinical implications.

---

---

Expected Outcome

Anticipated results include a clearer, more accurate low-dose CT image after processing, with evidence that diagnostic features are preserved and a recommended dose reduction range supported by data.