Breast Tomosynthesis Optimization Using AI-Assisted Image Reconstruction and Artifactual Noise Reduction

Project Abstract

Breast tomosynthesis presents a promising advancement over conventional 2D mammography by enabling improved lesion visibility through quasi-3D imaging, yet its clinical adoption is hampered by artifacts, noise, and reconstruction challenges that can affect diagnostic accuracy. This research investigates an AI-assisted image reconstruction framework designed to optimize image quality in breast tomosynthesis while robustly mitigating artifactual noise. The study integrates deep learning-based denoising, artifact suppression, and model-based iterative reconstruction to enhance lesion detectability, microcalcification visibility, and volumetric accuracy, with a focus on balancing denoising strength and?? edge preservation to maintain diagnostically relevant texture. A custom dataset comprising multi-view tomosynthesis projections, simulated and real-world acquisitions, with corresponding ground-truth references, is created to train and validate the proposed pipeline. The methodology comprises three core components (1) a physics-informed generative model that incorporates system geometry, dose constraints, and prior tissue statistical properties to drive accurate reconstruction; (2) an attention-guided multi-view fusion network that aggregates information across views to improve signal-to-noise ratio while preserving spatial resolution and mitigating view-dependent artifacts; and (3) a residual learning-based artifact reduction module trained to distinguish true anatomical features from reconstruction-induced distortions. The optimization objective combines fidelity to measured projections, perceptual quality metrics, and clinically relevant criteria such as lesion conspicuity, edge sharpness, and background parenchymal suppression. Comprehensive experiments compare the proposed method against standard weighted filtered back projection, model-based iterative reconstruction with conventional regularization, and state-of-the-art deep learning denoising approaches. Evaluation metrics include structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), lesion detectability indices, receiver operating characteristic (ROC) analysis, and reader studies with radiologists blinded to reconstruction technique. The results demonstrate statistically significant improvements in lesion conspicuity and calcification discrimination while achieving dose- and resolution-consistent reconstructions across diverse breast densities. The AI-assisted framework reduces artifactual streaks and secondary reflections commonly observed in high-contrast regions, leading to more accurate lesion extent estimation and reduced false-positive rates. Sensitivity analyses assess robustness to variations in acquisition geometry, noise levels, and dose, ensuring generalizability to clinical imaging workflows. Practical considerations address integration with existing tomosynthesis hardware, real-time or near-real-time inference, and software-hardware co-design to minimize latency without compromising image quality. The study also discusses potential regulatory and ethical implications, including data privacy, algorithm transparency, and the necessity for prospective multi-center trials to validate clinical impact. Overall, the research provides a comprehensive, scalable AI-enabled reconstruction strategy that enhances diagnostic performance in breast tomosynthesis by effectively suppressing artifactual noise while preserving critical structural detail, offering a viable pathway toward improved breast cancer screening and early detection.

Project Overview

What This Project Is About

---

A simple, reader-friendly exploration of improving breast tomosynthesis images by using artificial intelligence to reconstruct clearer images and reduce artifacts that can mislead diagnosis. The project looks at how AI algorithms can help combine multiple X-ray slices into a more accurate 3D picture of breast tissue while minimizing noise that is not part of the tissue.

---

---

The Problem It Addresses

---

Breast tomosynthesis can produce clearer, 3D views, but image artifacts and noise can obscure small details such as microcalcifications. Traditional reconstruction methods may require more radiation or longer processing. This project tackles how to make better images with smarter algorithms that maintain safety and speed.

---

---

Objectives of the Project

---

1. Understand the basics of breast tomosynthesis and AI reconstruction concepts.
2. Evaluate current reconstruction methods and identify artifact sources.
3. Develop or adapt AI models to improve image quality while keeping radiation dose safe.
4. Test artifact reduction techniques on simulated and real data.
5. Assess the impact on diagnostic features such as masses and calcifications.

---

---

What You Will Do Step by Step

---

1) Review foundational literature on tomosynthesis and AI image reconstruction. 2) Collect or generate sample breast imaging data with appropriate ethics approvals. 3) Implement AI-based reconstruction and artifact reduction methods. 4) Compare results against standard methods using objective metrics and radiologist feedback. 5) Analyze how dose and processing time change with AI methods. 6) Document limitations and potential clinical implications.

---

---

Expected Outcome

Anticipate clearer 3D breast images with fewer artifacts, improved detection of subtle features, and a demonstration that AI-assisted reconstruction can reduce the need for higher radiation doses or longer processing times while staying clinically acceptable.