Normal Pressure Hydrocephalus Classification using Weakly-Supervised Local Feature Extraction
9
Issued Date
2025-09-01
Resource Type
ISSN
00104825
eISSN
18790534
Scopus ID
2-s2.0-105011068694
Journal Title
Computers in Biology and Medicine
Volume
196
Rights Holder(s)
SCOPUS
Bibliographic Citation
Computers in Biology and Medicine Vol.196 (2025)
Suggested Citation
Supratak A., Kingchan S., Angkoondittaphong P., Nava-apisak P., Wongsripuemtet J., Noraset T., Kusakunniran W., Haddawy P., Songsaeng D. Normal Pressure Hydrocephalus Classification using Weakly-Supervised Local Feature Extraction. Computers in Biology and Medicine Vol.196 (2025). doi:10.1016/j.compbiomed.2025.110751 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/111392
Title
Normal Pressure Hydrocephalus Classification using Weakly-Supervised Local Feature Extraction
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
Normal Pressure Hydrocephalus (NPH) presents diagnostic challenges because its symptoms often overlap with other neurological conditions. A key radiological NPH indicator is ventricular cerebrospinal fluid (CSF) volume, assessed by neuroradiologists through brain imaging. However, expert availability is limited, highlighting the need for automated tools to assist in patient screening. While automated segmentation tools and pre-trained deep models have been used to estimate CSF volume for NPH diagnosis, they struggle to generalize to new datasets. This is partly due to their heavily reliance on prior knowledge, either through manually engineered features or the datasets used for pretraining segmentation models. Moreover, NPH classification models predominantly depend on global volume metrics, overlooking local CSF volume variations, which can result in suboptimal performance. In this paper, we introduce a new weak supervision method that can train a CSF segmentation model on a target dataset from scratch without needing costly segmentation annotations from experts. We also propose a local volumetric feature extraction algorithm that captures local differences in CSF volumes across different brain partitions along the axial orientation, providing richer information beyond the global metrics. Our weakly-supervised CSF segmentation model, combined with local volumetric features, was evaluated on non-contrast CT scans of 105 NPH and 112 non-NPH patients. The results show that our approach outperforms existing segmentation methods in NPH classification performance (ACC = 0.88, Sen = 0.97, Spec = 0.79, F1 = 0.89, AU-ROC = 0.91). Our model also demonstrates superior performance in screening patients at risk of having NPH compared to the visual-based evaluation of neuroradiologists.