Related Concept Videos
You might also read
Related Articles
Articles linked to this work by shared authors, journal, and citation graph.
You might also read
Articles linked to this work by shared authors, journal, and citation graph.
Updated: Oct 5, 2025

Three-Dimensional Imaging of Tumor-Bearing Tissue Using the Iterative Bleaching Extends Multiplexity Approach
Published on: April 25, 2025
Yunsu Choi1, Minah Han1, Hanjoo Jang1
1School of Integrated Technology and Yonsei Institute of Convergence Technology, Yonsei University, Incheon, South Korea.
This study introduces a two-step artificial intelligence method to remove blur from 3D breast images. By processing data in a coarse-to-fine sequence, the technique improves image clarity and lesion detection across multiple viewing angles.
Area of Science:
Background:
Limited projection angles in breast imaging systems often result in significant reconstruction artifacts. These distortions frequently obscure anatomical details and hinder the identification of potential lesions. Prior research has shown that standard reconstruction techniques struggle to maintain high fidelity across all spatial planes. That uncertainty drove the need for advanced computational solutions to enhance diagnostic precision. No prior work had resolved the challenge of multi-directional blurring using a sequential learning framework. Existing approaches often fail to address the complex nature of these artifacts in three-dimensional volumes. This gap motivated the development of a specialized model to restore missing frequency information. The current landscape of medical imaging requires more robust tools to support clinical interpretation.
Purpose Of The Study:
The study aims to develop a two-phase learning approach for compensating blurring artifacts in breast imaging. This research addresses the severe degradation caused by limited-angle projection data in reconstruction systems. The authors seek to improve the detection performance of lesions by enhancing image clarity. They propose a coarse-to-fine strategy to mitigate blurring along axial, coronal, and sagittal planes. The motivation stems from the difficulty of identifying abnormalities in images with significant artifacts. By employing a convolutional neural network, the team intends to restore missing frequency information. This work explores how different loss functions influence the overall deblurring performance of the model. The primary goal is to provide a robust solution for improving the diagnostic utility of tomosynthesis volumes.
Main Methods:
Review Approach involved designing a two-phase learning framework for artifact compensation. The researchers implemented a convolutional neural network to process image volumes in a coarse-to-fine sequence. Phase 1 focused on three-dimensional restoration, while Phase 2 addressed additional two-dimensional refinement. The team evaluated various loss functions, including pixel-based, adversarial-based, and perception-based criteria. They tested the model's ability to mitigate blurring across axial, coronal, and sagittal planes. Quantitative validation relied on comparing the processed outputs against the original reconstructed images. The investigators assessed the restoration of missing frequency components through the sequential steps. This systematic evaluation ensured that the model effectively handled the complexities of limited-angle projection data.
Main Results:
Key Findings From the Literature show the proposed method significantly reduces blurring artifacts. The mean squared error of the image decreased by 82.8% compared to the original data. Furthermore, the root mean squared errors of the gradient of the image dropped by 44.9%. The contrast-to-noise ratio in the in-focus plane increased by 183.4%. These improvements demonstrate the effectiveness of the coarse-to-fine learning strategy. The researchers verified that the model sequentially restores missing frequency components during the two-phase process. The results indicate a substantial reduction in artifacts across all viewing directions. This performance supports the utility of the model for enhancing diagnostic image quality.
Conclusions:
The authors propose that their sequential learning framework effectively mitigates blurring across multiple spatial planes. Synthesis and implications suggest that this coarse-to-fine strategy enhances the visibility of diagnostic features. The researchers claim that their model successfully restores missing frequency components through the two-phase processing architecture. Evidence indicates that the approach significantly reduces artifacts compared to standard reconstruction methods. The study demonstrates that the technique improves the contrast-to-noise ratio in the in-focus plane. These findings imply that the method could bolster the detection performance of lesions in clinical settings. The authors conclude that their dual-phase architecture provides a robust solution for correcting image degradation. This work highlights the potential of deep learning to improve the quality of tomosynthesis imaging.
The researchers propose a two-phase architecture where Phase 1 performs 3D deblurring, followed by Phase 2 for 2D refinement. This sequential process restores missing frequency components, leading to an 82.8% reduction in mean squared error and a 183.4% increase in the contrast-to-noise ratio.
The model utilizes a convolutional neural network. This architecture is divided into two distinct submodels to handle the coarse-to-fine restoration of image data across axial, coronal, and sagittal planes.
The authors evaluated pixel-based, adversarial-based, and perception-based loss functions. These different mathematical constraints were necessary to optimize the network's ability to minimize blurring artifacts while preserving structural integrity during the reconstruction process.
The study employs projection data acquired from limited angles. This input is processed through the network to compensate for artifacts that typically degrade the quality of reconstructed volumes.
The researchers measured the mean squared error of the image and the root mean squared errors of the gradient. These metrics showed decreases of 82.8% and 44.9%, respectively, confirming the effectiveness of the proposed restoration technique.
The authors suggest that their method improves lesion detection performance. By reducing blurring artifacts in the in-focus plane and across other spatial orientations, the model facilitates more accurate identification of clinical findings compared to unprocessed images.