Page 109 - AIH-2-3

Artificial Intelligence in Health Bone suppression utility for chest diagnosis

need for further validation, as AI model performance is In addition, we employed Grad-CAM to visualize the
often influenced by biases in the training data. Specifically, rationale for the model’s decisions. We confirmed that the
the small dataset size may have led to imbalances in disease bone suppression techniques effectively directed the model’s
severity, patient age distribution, and gender ratio. In fact, focus to the inner lung field. The effect could enhance the
the mean label score of 0.380 and SD of 0.260 suggest explainability and interpretability of the models, thereby
that most of the images are associated with lower severity increasing confidence in their predictions. However, even
scores, potentially limiting the model’s performance and when the activated areas of the models were focused within
generalizability. Consequently, the effectiveness of bone the lung fields, the predicted scores did not consistently
suppression for severity assessment may not be applicable align with the true scores. Consequently, improving
to all patient populations or clinical settings. Furthermore, prediction accuracy while fostering explainability remains
artificially generated images can exhibit undesirable a challenge for achieving broader acceptance as a reliable
artifacts or implausible shadows, even when produced diagnostic support tool.
55
by cutting-edge models. Therefore, the quality of the In summary, this study improved the clinical applicability
bone-suppressed images should be carefully inspected of the AI-based bone suppression system by eliminating
by experts, such as radiologists and respiratory medicine the need for subtraction processing. The updated system
physicians, to identify potential artifacts and ensure that enhanced the visibility of lung abnormalities, leading to more
the visibility of lesions is preserved.
accurate predictions of pneumonia severity, as demonstrated
We applied the updated bone suppression model as a by statistically significant improvements in the linear
pre-processing step in COVID-19 severity assessment using correlation between predicted severity scores and actual
regression models. By utilizing the bone-suppressed images, labels. These findings underscore the utility of AI-driven
we observed statistically significant improvements in linear bone suppression in CXR, particularly for regression tasks
correlations between predicted scores and true labels across related to disease severity, progression, and recurrence risk.
various regression models, including classical models such Furthermore, the bone suppression techniques guided the
as ResNet18 and the newer model RegNetY-120. Notably, prediction model to concentrate on the inner lung fields,
the highest PCC of 0.895 exceeded the performance suggesting potential improvements in the reliability of clinical
reported in related studies. 32,34 However, the DenseNet assessments. The application of bone suppression in assessing
model showed no significant differences in performance COVID-19 severity could optimize patient monitoring and
with and without bone suppression. We acknowledge that, healthcare resource allocation. In addition, this advancement
again, the small dataset size remains a significant limitation, has the potential to elevate the diagnostic accuracy of CXR,
potentially affecting the generalizability of this study. providing valuable tools to overcome existing limitations,
Furthermore, the variability in performance across different such as inferior contrast resolution and the superimposition
regression models warrants further investigation. In future of anatomical structures compared to CT.
work, we plan to confirm the effectiveness and robustness of
bone suppression techniques using larger and more diverse 5. Conclusion
datasets and a broader range of prediction models. This study successfully developed and validated an
The clinical implications of this study are promising. AI-based bone suppression model for CXR, which
The proposed bone suppression system has the potential to effectively removes bone structures while highlighting
enhance the diagnostic performance of CXR by improving lung abnormalities. The model not only improves image
the visibility of abnormalities and enabling more accurate quality but also streamlines the entire image processing
disease assessment. Beyond its established effectiveness workflow, increasing its clinical practicality. We applied
in classification tasks, bone suppression techniques could the bone suppression model as a pre-processing step to
facilitate regression-based evaluations of disease severity, facilitate more accurate predictions of COVID-19 severity.
progression, and recurrence risk. Integrating this system The findings demonstrate the potential of bone suppression
into diagnostic image viewing software would allow techniques in assessing various pulmonary conditions,
radiologists and physicians to access bone-suppressed particularly through regression analyses. Future research
images as additional clinical information. Moreover, as should focus on validating bone suppression techniques
demonstrated in our assessment of COVID-19 severity, with larger and more diverse datasets, as well as exploring
these techniques could serve as a pre-processing tool a range of prediction models. In addition, addressing
for computer-aided diagnosis systems. This integration potential biases in AI outputs and enhancing the model’s
may help reduce the discrepancy in diagnostic accuracy explainability will be essential for ensuring its reliable
between CXR and CT. integration into routine clinical practice.

Volume 2 Issue 3 (2025) 103 doi: 10.36922/aih.5608