Automated Idiopathic Normal Pressure Hydrocephalus Diagnosis via Artificial Intelligence–Based 3D T1 MRI Volumetric Analysis

Joonhyung Lee; Dana Kim; Chong Hyun Suh; Suyoung Yun; Kyu Sung Choi; Seungjun Lee; Wooseok Jung; Jinyoung Kim; Hwon Heo; Woo Hyun Shim; Sungyang Jo; Sun Ju Chung; Jae-Sung Lim; Ho Sung Kim; Sang Joon Kim; Jae-Hong Lee

doi:10.3174/ajnr.A8489

Compose eLetter

Title *

Author Information

Contributors
First Name and Middle Initial * First or given name, e.g. 'Peter'. Last Name * Your last, or family, name, e.g. 'MacMoody'. Email Address * Your email address, e.g. higgs-boson@gmail.com Occupation * Your role and/or occupation, e.g. 'Orthopedic Surgeon'. Affiliation * Your organization or institution (if applicable), e.g. 'Royal Free Hospital'.

Statement of Competing Interests

Competing interests? *

Yes

Please describe the competing interests

CAPTCHA

Please verify that you are a real person

Vertical Tabs

Jump to comment:

RE:
Chong Hyun Suh and Su Young Yun

Published on: 06 April 2025
Advancements in AI for Idiopathic Normal Pressure Hydrocephalus Diagnosis
Mingjia Li, Yuwei Dai, Li Yang and Harrison X. Bai

Published on: 17 March 2025

Published on: (6 April 2025)
Page navigation anchor for RE:
RE:
Chong Hyun Suh, M.D., Ph.D., University of Ulsan College of Medicine Asan Medical Center Seoul, Republic of Korea
Other Contributors:
Su Young Yun, M.D.
We thank Mingjia et. Al for valuable feedback on our article. We would also like to express our gratitude to AJNR for providing the opportunity to response.
First, as mentioned in the limitations of our study, it did not reflect the real-world prevalence of iNPH. Since iNPH is a rare disease, we had to use an imbalanced dataset to include a sufficient number of iNPH patients. However, we believe that this study demonstrates the robustness of the AI model through results obtained using cross-validation and an unseen test dataset.
Secondly, it would be a valuable research direction to explore alternative AI approaches and compare them with the AI model we developed. Further research will be needed in this area.
Thirdly, regarding the point mentioned about the lack of external validation across diverse scan parameters and population characteristics, which questions the model’s adaptability to real-world clinical environments, as stated in the Methods section, we conducted cross-validation and provided results from both the training dataset and the unseen test dataset. The training dataset and unseen test dataset come from different hospitals, and the MRI vendors used are also different. Moreover, since the unseen test dataset includes data from various MRI machines, it demonstrates the robustness of the AI model in handling real-world heterogeneity.
Additionally, while the AI algorithm lacks interpretability due to its black-box nature, this AI model m...
Show More

We thank Mingjia et. Al for valuable feedback on our article. We would also like to express our gratitude to AJNR for providing the opportunity to response.
First, as mentioned in the limitations of our study, it did not reflect the real-world prevalence of iNPH. Since iNPH is a rare disease, we had to use an imbalanced dataset to include a sufficient number of iNPH patients. However, we believe that this study demonstrates the robustness of the AI model through results obtained using cross-validation and an unseen test dataset.
Secondly, it would be a valuable research direction to explore alternative AI approaches and compare them with the AI model we developed. Further research will be needed in this area.
Thirdly, regarding the point mentioned about the lack of external validation across diverse scan parameters and population characteristics, which questions the model’s adaptability to real-world clinical environments, as stated in the Methods section, we conducted cross-validation and provided results from both the training dataset and the unseen test dataset. The training dataset and unseen test dataset come from different hospitals, and the MRI vendors used are also different. Moreover, since the unseen test dataset includes data from various MRI machines, it demonstrates the robustness of the AI model in handling real-world heterogeneity.
Additionally, while the AI algorithm lacks interpretability due to its black-box nature, this AI model mitigates that weakness by using a decision tree. Decision tree is locally interpretable Model-Agnostic Explanations, which improves interpretability.1 This approach utilizes a logical structure similar to how neuroradiologists assess iNPH. This can be more easily understood by referring to Figure 1. Because of this, the model may have lower discriminative ability compared to other AI tools, but it offers higher interpretability. Lastly, while we believe that we have validated the model on a heterogeneous dataset, AI cannot be fully trusted, and thus, it cannot replace radiologists. However, we believe that the development of such models can assist neurologist and neuroradiologist in reducing underdiagnosis and misdiagnosis, ultimately benefiting the clinical environment.
Reference
1. Ennab, Mohammad, and Hamid Mcheick. Designing an interpretability-based model to explain the artificial intelligence algorithms in healthcare. Diagnostics 2022;12(7):1557.
Show Less

Competing Interests: None declared.
Published on: (17 March 2025)
Page navigation anchor for Advancements in AI for Idiopathic Normal Pressure Hydrocephalus Diagnosis
Advancements in AI for Idiopathic Normal Pressure Hydrocephalus Diagnosis
Mingjia Li, Student, Johns Hopkins School of Medicine, Department of Radiology and Radiological Sciences, Baltimore, MD, US
Other Contributors:
Yuwei Dai, Research Fellow

Li Yang, Associate Professor

Harrison X. Bai, Associate Professor of Radiology and Radiological Science
Idiopathic normal pressure hydrocephalus (iNPH), a reversible form of dementia accounting for an estimated 1.6–5.4% of dementia cases, is frequently underdiagnosed and misdiagnosed due to its symptomatic overlap with conditions like Alzheimer’s (AD) and Parkinson’s (PD). [1] Recent advances in medical imaging and artificial intelligence (AI), however, offer promising solutions, including a study by Lee et al. that uses AI-driven 3D T1-weighted MRI volumetric analysis to identify key brain features associated with iNPH and automate measuring biomarkers. [2] Their model demonstrated robust diagnostic performance, achieving an area under the curve (AUC) of 0.956 for high-convexity tightness and 0.830 for Sylvian fissure enlargement. Additionally, cross-validation and unseen test set yielded AUCs of 0.983 and 0.936, respectively. Such performance metrics underscore how AI-driven tools can enhance diagnostic accuracy and streamline workflows in clinical settings by enabling rapid, automated analysis of neuroimaging data, addressing the need for large-scale screenings for targeted care and effective intervention.

However, despite its advancements, Lee et al.’s study has several limitations that require careful consideration. First, the preselected cohort of iNPH, PD, AD, and healthy controls (HC) may not reflect the real-world prevalence of these diseases, introducing selection bias and raising concerns about the model’s generalizability. For instance, the cohort’s iNPH...
Show More

Idiopathic normal pressure hydrocephalus (iNPH), a reversible form of dementia accounting for an estimated 1.6–5.4% of dementia cases, is frequently underdiagnosed and misdiagnosed due to its symptomatic overlap with conditions like Alzheimer’s (AD) and Parkinson’s (PD). [1] Recent advances in medical imaging and artificial intelligence (AI), however, offer promising solutions, including a study by Lee et al. that uses AI-driven 3D T1-weighted MRI volumetric analysis to identify key brain features associated with iNPH and automate measuring biomarkers. [2] Their model demonstrated robust diagnostic performance, achieving an area under the curve (AUC) of 0.956 for high-convexity tightness and 0.830 for Sylvian fissure enlargement. Additionally, cross-validation and unseen test set yielded AUCs of 0.983 and 0.936, respectively. Such performance metrics underscore how AI-driven tools can enhance diagnostic accuracy and streamline workflows in clinical settings by enabling rapid, automated analysis of neuroimaging data, addressing the need for large-scale screenings for targeted care and effective intervention.

However, despite its advancements, Lee et al.’s study has several limitations that require careful consideration. First, the preselected cohort of iNPH, PD, AD, and healthy controls (HC) may not reflect the real-world prevalence of these diseases, introducing selection bias and raising concerns about the model’s generalizability. For instance, the cohort’s iNPH prevalence (24.6% of 452 patients) far exceeds real-world estimates. Second, the reliance on a single machine learning classifier, XGBoost, without comparison to alternative AI approaches, such as deep learning or vision-language models, limits insights into whether new methodologies could enhance performance, especially with AI’s rapid evolution in medical imaging. Third, the lack of external validation across diverse scan parameters and population characteristics questions the model’s adaptability to real-world clinical environments, where variability in equipment and patient demographics is common. These methodological constraints highlight the challenges of translating AI innovations into practice.

Beyond technical limitations, broader hurdles persist in integrating AI into healthcare. Algorithm interpretability remains a critical barrier, as clinicians require transparent decision-making processes to trust AI outputs. Data generalizability is another concern—models must perform consistently across heterogeneous datasets to avoid biases. Also, AI should complement—not replace—clinical judgment, rather serving as a screening tool to flag potential cases for specialist review. While Lee et al.’s model shows high sensitivity, validating its utility against diagnoses by experienced radiologists and epidemiological benchmarks ensures reliability.

Overall, Lee et al. represents a promising step toward advancing iNPH diagnosis. Al’s ability to streamline workflows and reduce delays could address underdiagnosis and misdiagnosis. Future efforts must prioritize multi-center trials, diverse AI comparisons, and rigorous validation. As AI evolves, clinician-researcher collaboration must guide development to align tools with real-world needs, enabling earlier interventions and improved patient outcomes.

References
1. Bendella Z, Purrer V, Haase R, et al. Brain and ventricle volume alterations in idiopathic normal pressure hydrocephalus determined by artificial intelligence-based MRI volumetry. Diagnostics (Basel) 2024;14(13):1422. DOI: https://doi.org/10.3390/diagnostics14131422
2. Lee J, Kim D, Suh CH, et al. Automated idiopathic normal pressure hydrocephalus diagnosis via artificial intelligence-based 3D T1 MRI volumetric analysis. AJNR Am J Neuroradiol 2025;46(1):33-40. DOI: https://doi.org/10.3174/ajnr.A8489
Show Less

Competing Interests: None declared.