University of Crete (UOC)
Non-invasive estimation of intracranial pressure from computed tomography images using artificial intelligence methods
Abstract
dc:descriptionThis study introduces a novel Deep Learning based approach in order to estimate Intracranial Pressure (ICP) non-invasively. It uses standard CT brain scans in combination with patient data (clinical and demographic). Elevated Intracranial Pressure is a critical condition in neurological cases like Traumatic Brain Injury and hydrocephalus, influencing cerebral perfusion. Potentially it can cause irreversible brain damage or even death. While invasive methods such as external ventricular drainage and intraparenchymal probes are considered the most accurate way to monitor Intracranial Pressure, they carry considerable dangers, such as hemorrhage, infection, high cost, and the requirement for expert neurosurgical intervention. This highlights the urgent need for a reliable, accurate, and rapid non-invasive monitoring technique. This research was conducted as a retrospective observational study that analyzed a dataset of 578 patient cases from PAGNI hospital, gathered from 2012 to 2023. The research included patients who had both a CT scan and an associated ICP measurement taken within a strict ±1-day window. ICP was mainly measured using fiberoptic intraparenchymal catheters or intraventricular catheters. Patients were then classified into two groups based on a clinically relevant threshold of 15 mmHg: 140 pathological cases (ICP ≥ 15 mmHg) and 438 normal cases (ICP < 15 mmHg). A comprehensive preprocessing pipeline was applied to the data, including anonymization, Hounsfield Unit conversion, and windowing for better tissue contrast. For consistency, images were resampled to a fixed dimension (120 × 128 × 128), then normalized, and finally skull segmented in order to remove irrelevant regions. A significant innovation of this study was the way that demographic data (age, sex, Glasgow Coma Scale scores) were integrated. This was achieved by encoding and embedding these data into synthetic 2D image slices, which were then combined with the CT volumes, creating a single, unified input for the Deep Learning models. Four distinct Deep Learning models were investigated: a custom-designed 3D Convolutional Neural Network, and three 3D adaptations of ImageNet (DenseNet201, ResNet101, and MobileNetV2) that were initially pretrained. Training strategies included binary cross-entropy loss, the Adam optimizer, ReLU activation for hidden layers, and sigmoid for the output. Also, dropout and L2 regularization were implemented in order to mitigate overfitting. To manage the imbalanced dataset 1:3 pathological-to-normal ratio), data augmentation (random rotations, Gaussian blurring for minority class) and class weighting were used. The models were assessed by using Area Under the Receiver Operating Characteristic Curve (AUC), Area Under the Precision-Recall Curve (PR AUC), F1 score, recall (Sensitivity), and precision. The MobileNetV2 3D model was emerged clearly as the top-performing model, especially when demographic data were included. It achieved an impressive AUC of 0.883, a PR AUC of 0.694, and a recall of 0.818. The inclusion of demographic data significantly boosted its performance, leading to a 9.9% increase in AUC and a 21.1% rise in PR AUC for MobileNetV2 3D, highlighting the benefits of combining different data types. DenseNet201 3D also performed well, regardless of whether demographic data was included. In contrast, ResNet101 3D consistently failed to distinguish between classes, performing no better than random chance with an AUC of 0.500.Class Activation Maps were used to provide visual explanations of the models' decision-making process. MobileNetV2 3D heatmaps clearly showed concentrated activity around important brain areas, like the ventricular system and periventricular regions. This aligns with known signs of high Intracranial Pressure. This contrasts with the diffuse and non-specific activations from the poorly performing ResNet101 3D, emphasizing how important it is for a model to be interpretable for clinical trust.The study's findings proves that Artificial Intelligence can classify Intracranial Pressure using routine CT scans, offering by this way a practical and scalable solution, comparable to other non-invasive methods. The models’ high recall values make them suitable for clinical screening, where sensitivity in detecting pathological cases is critical. Despite the limitations of the study, such as the relatively small dataset, the class imbalance, temporal inconsistencies, and the lack of external validation, it lays a strong foundation for future research. Future directions focus on gathering larger datasets from multiple centers, looking into multi-class or regression approaches, adding time-based data, improving explainability, and further integrating multi-modal clinical information.
Degree
thesis:*- Grantor dc:publisher
- University of Crete (UOC)
- Year dc:date
- 2025
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Theodoropoulos, Dimitrios
- Θεοδωρόπουλος, Δημήτριος
Subjects
dc:subject × 12Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/59930