Optimization of radiation dose reduction in pediatric chest radiography using deep learning–driven automated exposure control

Project Abstract

This study presents a novel framework for reducing radiation dose in pediatric chest radiography through a deep learning–driven automated exposure control (AEC) system that dynamically adapts exposure parameters to patient size, anatomy, and clinical indication. We design and train a multi-tasks neural network that estimates optimal kilovoltage peak (kVp), milliampere-seconds (mAs), and exposure time from low-dose pre-scan inputs, patient demographics, and real-time feedback from initial scout views. The framework combines a reinforcement learning agent with a supervised regression model the former optimizes exposure settings across views to minimize dose while preserving diagnostic image quality, and the latter provides robust initial parameter estimation to accelerate convergence. A physics-informed loss function integrates image quality metrics (contrast-to-noise ratio, edge sharpness, structural similarity) with dose constraints derived from pediatric radiology guidelines, ensuring adherence to ALARA principles. To ensure clinical reliability, we develop a hybrid calibration strategy that leverages phantom-based dose measurements and retrospective pediatric chest radiographs annotated by radiologists for diagnostically acceptable thresholds. Our dataset comprises multi-institutional pediatric chest radiographs with corresponding dose records and acquisition settings, augmented with synthetic variations to simulate a wide range of body habitus and clinical scenarios. The proposed AEC system is implemented as an end-to-end pipeline a pre-scan assessment module estimates patient size and projection-specific attenuation; an exposure decision engine proposes initial settings; a real-time feedback loop refines exposure during the acquisition based on interim low-dose priors and histogram analysis; and a post-processing module applies denoising and perceptual optimizations to further enhance image quality without increasing dose. We evaluate performance against standard fixed-exposure protocols and adaptive manual protocols across multiple endpoints dose saved (percent reduction in effective dose and entrance skin dose), diagnostic quality (radiologist scoring and task-based image quality metrics), and robustness (failure modes under motion, obesity, and unusual anatomy). Our results demonstrate a statistically significant reduction in mean effective dose by [specific percentage range to be determined from results], while maintaining diagnostic confidence scores within inter-observer variability limits. The system shows strong generalization across institutions, scanner models, and projection angles, with minimal need for manual parameter tuning. Sensitivity analyses reveal dose-quality trade-offs under varying motion and noise conditions, informing clinically acceptable operating envelopes. Additionally, the framework provides explainable AI components by visualizing attention maps over anatomical regions contributing to exposure decisions, and by correlating parameter choices with predicted image quality outcomes. The study discusses potential integration into radiography workflows, the implications for pediatric patient safety, regulatory considerations, and pathways for prospective clinical validation. Limitations include the need for large-scale multicenter validation, potential biases in dataset composition, and the challenge of standardizing image quality metrics across platforms. Future work will explore federated learning to preserve patient privacy, real-time hardware synchronization for dose modulation, and extension to other pediatric imaging modalities.

Project Overview

What This Project Is About

---

A plain-language overview of the topic and what the project investigates.

---

---

The Problem It Addresses

---

What problem or gap this project tackles and why it matters to the field or society.

---

---

Objectives of the Project

---

1. Identify how radiation dose varies in pediatric chest X-rays and what factors influence it.
2. Explore how automated exposure control could tailor doses to individual patients.
3. Evaluate how deep learning can help predict safe exposure settings without compromising image quality.
4. Propose a workflow that combines dose reduction with reliable diagnostic results.

---

---

What You Will Do Step by Step

---

1) Review basics of radiography and dose concepts in children and gather relevant guidelines.

2) Collect or simulate pediatric chest X-ray data with varying exposure settings.

3) Develop a simple deep learning model to suggest exposure settings based on patient factors.

4) Test whether the model maintains image quality while reducing dose using objective metrics.

5) Compare automated exposure results with standard practices and assess practicality.

---

---

---

---

Expected Outcome

It is expected to produce a framework where a lightweight model assists clinicians in selecting lower, safe radiation doses that still yield usable diagnostic images, along with a report on feasibility for clinical integration.