Page 103 - AIH-2-3

Artificial Intelligence in Health Bone suppression utility for chest diagnosis

involvement and the degree of opacity, achieving Pearson Healthcare, Chicago, IL, USA) at Kitasato University
correlation coefficients (PCCs) of 0.80 and 0.78 for these Hospital (Sagamihara City, Japan), to develop a bone
tasks, respectively. Signoroni et al. introduced BS-Net, suppression model. Most of these patients had pulmonary
34
an end-to-end architecture, to segment, align, and inflammatory diseases or pulmonary mass lesions.
quantify lung compromise based on the Brixia score. 35,36 The detector specifications are detailed in our previous
The performance of BS-Net was evaluated not only for work. Radiography was performed with tube voltages of
22
classification tasks but also for regression tasks using 130 kV for high-energy images and 60 kV for low-energy
linear regression of the Brixia score, with the highest PCC images. The system produces bone-suppressed and bone-
reaching 0.85. These studies underscore the potential of AI enhanced images, along with standard chest radiographs
approaches in evaluating COVID-19 severity. for presentation, all with a resolution of 3524 × 4288 pixels
Moreover, transparency, explainability, and and 13-bit contrast, from the raw data of the high- and
interpretability are critical components of AI, especially low-energy images. For training, we utilized 480 pairs of
37
in medical applications. Understanding why and how standard and bone-suppressed radiographs, while 120
a model derives a particular decision is essential for pairs were reserved for testing.
ensuring clinical accountability and building confidence. 2.1.2. Data pre-processing
Gradient-weighted class activation mapping (Grad-CAM)
is a widely used technique to visualize decision-making To prepare the dataset for model training, we first cropped
processes by highlighting image regions that contribute the standard and bone-suppressed radiographs to extract
to the model’s output. By generating heatmaps based on regions of interest (ROIs) centered on the lung area. The
38
41
gradients from the final convolutional layer, Grad-CAM lung regions were identified using a pre-trained U-Net
offers explainable insights to support clinical decision- model, which segments chest radiographs into the lung,
making. 39,40 Talaat et al. integrated Grad-CAM into a heart, other anatomical areas, and background, assigning
40
breast cancer classification model, providing radiologists pixel values of 255, 85, 170, and 0, respectively, in 8-bit
with valuable insights into the model’s decision-making contrast. The U-Net architecture employed consisted of
process and fostering trust in the AI system. In this study, five depths, incorporating an input layer, five encoder
Grad-CAM is used to validate the explainability and layers, five decoder layers, and an output layer.
interpretability of COVID-19 severity prediction models. For training the U-Net model, we utilized all 247 chest
Our work integrates AI-based bone suppression pre- radiographs from the Japanese Society of Radiological
processing into regression models to assess COVID-19 severity Technology database, along with their corresponding
using CXR. The primary aim is to expand the applications of segmented labels. The U-Net model was trained for up to
42
AI-based bone suppression techniques, verifying their utility 100 epochs using the RMSprop optimizer, with a learning rate
in severity assessment. By improving the accuracy of severity of 0.0001, a weight decay of 1 × 10 , and a momentum of 0.9.
−8
predictions, this approach could enhance patient monitoring After training, the U-Net model was applied to identify
and optimize healthcare resource allocation, particularly in the lung regions in the standard radiographs collected at
resource-limited settings. Our findings may also validate Kitasato University Hospital, which had been converted
the applications of AI-based bone suppression in regression to 8-bit contrast in advance. These identified lung regions
tasks for chest image diagnosis. Moreover, this study seeks were then cropped from both the standard and bone-
to bridge the gap between the present limitations of CXR suppressed radiographs. Finally, the cropped images were
and the superior sensitivity of CT, ultimately contributing to
more efficient and scalable diagnostic tools for COVID-19 resized to 1024 × 1024 pixels to standardize the input size
and other pulmonary diseases. for the subsequent training of the bone suppression model.
2. Data and methods 2.1.3. Bone suppression network architecture and
training settings
In this section, we explain the development of the bone We employed the pix2pix 43,44 network to generate
suppression model, followed by the method for assessing virtually bone-suppressed images from the standard
COVID-19 severity. chest radiographs. Figure 1 illustrates a flowchart of the
2.1. Bone suppression model bone suppression and pre-processing steps. The network
architecture follows the design proposed by Isora et al.,
43
2.1.1. Data collection as described in our previous work, with modifications
22
We collected chest radiographs from 600 patients made to the resolution of the generator and discriminator
using a dual-shot DES system (Discovery XR656, GE to handle 1024 × 1024 resolution images.

Volume 2 Issue 3 (2025) 97 doi: 10.36922/aih.5608