Page 102 - AIH-2-3

Artificial Intelligence in Health Bone suppression utility for chest diagnosis

mild symptoms. Mild cases are often self-limiting and do bone edge detection to generate both bone-enhanced and
5
not require hospitalization, whereas severe cases necessitate bone-suppressed images, demonstrating the potential for
admission to the intensive care unit; mechanical ventilation, improved pneumonia diagnosis in children.
including extracorporeal membrane oxygenation; and We also previously developed an AI-DES system that
treatment with anti-inflammatory agents. An early successfully generated bone-suppressed images; however,
6
and accurate diagnosis, along with a reliable severity its clinical applicability was limited by the need for raw
assessment, is critical for effective patient management and high- and low-energy images, which are often unavailable,
preventing the overburdening of healthcare facilities. and by labor-intensive weighted subtraction processing.
22
Chest radiography (CXR) is widely used for initial To overcome these limitations, we updated the system to
evaluations in medical emergencies due to its low cost, generate bone-suppressed images directly from routine
rapid examination, and widespread availability. While chest radiographs in this study. This improvement enhances
7
CXR is useful for detecting COVID-19, particularly during the system’s versatility and clinical utility by streamlining
pandemic situations, its sensitivity is lower than that of the entire image-processing workflow.
chest computed tomography (CT). Typical imaging Several studies have demonstrated the advantages of
8,9
patterns observed in COVID-19 patients include bilateral, AI-based bone suppression techniques in classification tasks
peripheral, and basal-predominant ground-glass opacities related to various lung diseases, including COVID-19. 23-29
without pleural effusion. 8,10,11 Although CT is generally In particular, Rani et al. proposed a model that preserves
23
considered more reliable for identifying such subtle contrast spatial features, suggesting that effectively pre-processed
lesions, the potential of image pre-processing techniques radiographs could enhance diagnostic performance. The
9
to enhance COVID-19 detection using CXR has been authors further reported that integrating AI-based bone
extensively explored. 12-14 For instance, Sharifrazi et al. suppression with pre-processing techniques, such as lung
12
demonstrated that Sobel filter pre-processing improves segmentation and augmentation, significantly improves
the performance of convolutional neural network (CNN) the classification accuracy of pneumonia, including
models for detecting COVID-19 using CXR. Similarly, COVID-19. Lam et al. reported that bone-suppressed
24
25
Arias-Garzón et al. reported that pre-processing with images, generated using a CNN-based model proposed by
13
segmentation techniques leads to a more accurate and Rajaraman et al., significantly increased the area under
26
reliable COVID-19 classification. Avolio et al. further the receiver operating characteristic curve for COVID-
14
suggested that pre-processing methods could benefit 19 classification tasks, compared to standard radiographs
not only CNN-based approaches but also other machine using a modified VGG16 model. Similarly, Xu et al.
30
learning techniques, such as multiple-instance learning.
developed a CNN-based rib removal model, SADXNet,
Furthermore, Takaki et al. demonstrated that bone which showed superior performance in lung nodule
15
suppression processing through temporal subtraction detection, lung anomaly classification, and localization
significantly increased radiologists’ sensitivity and reduced tasks. These findings highlight the importance of bone
their false-positive rates in detecting pulmonary lesions suppression in enhancing the accuracy of lung disease
on CXR. In contrast, van der Heyden suggested that classification.
16
dual-energy subtraction (DES) technology enhances the Given these promising results, we hypothesize that bone
diagnostic accuracy of CXR by eliminating bone structures suppression techniques may also be effective in regression
and improving soft-tissue visualization. Building on these tasks, such as assessing disease severity, predicting the risk of
conventional bone suppression methods, advancements progression, and estimating patient prognosis. Regression
in artificial intelligence (AI) have introduced innovative models, which predict continuous values by identifying
alternatives for enhancing bone suppression. 17-30 Unlike subtle variations, can be more complex and challenging
conventional DES systems, which rely on dual X-ray to construct than classification models, particularly when
exposure and specialized detectors equipped with copper dealing with high-dimensional or imbalanced data. This
31
plates, AI-based DES (AI-DES) approaches improve
accessibility while offering additional advantages, such as study explores the utility of AI-based bone suppression in
reduced motion artifacts and improved noise characteristics. COVID-19 severity assessment using CXR, comparing the
A representative study employed a generative adversarial performance of regression models with and without bone
network (GAN)-based model to suppress bone structures suppression to validate its effectiveness.
in CXR without the need for labeled data, utilizing Recent studies have proposed AI-based prediction
digitally reconstructed radiographs from CT data. Recent models for COVID-19 severity assessment. 32-34 Cohen
18
advancements have further extended these applications et al. 32,33 used DenseNet-based regression models to
to pediatric imaging. 19,20 For example, Xie et al. utilized evaluate COVID-19 severity based on the extent of lung
20

Volume 2 Issue 3 (2025) 96 doi: 10.36922/aih.5608