Feasibility comparison of deep learning image regressions to estimate intracranial pressure from cranial computed tomography in hydrocephalus

INTRODUCTION

One of the neurosurgical disorders that cause functional disability globally is hydrocephalus (HCP), especially in low- and middle-income nations.^[1,2] It results from disturbances in cerebrospinal fluid (CSF) dynamics, leading to ventricular dilatation and elevated intracranial pressure (ICP), both of which are related to unfavorable outcomes.^[3,4] Therefore, external ventricular drainage (EVD) and other CSF diversion procedures remain standard interventions to control ICP.

Elevated ICP is a key prognostic factor across neurosurgical conditions. In traumatic brain injury, ICP >20 mmHg has been associated with mortality and reduced cerebral perfusion pressure,^[5] while levels above 25–30 mmHg predict unfavorable outcomes in subarachnoid and intraventricular hemorrhage.^[6,7] Accurate prediction of ICP is therefore critical for timely intervention and prognostication.

Machine learning (ML) and deep learning (DL) approaches have increasingly been applied for ICP prediction. Prior studies have reported strong performance of random forests and logistic regression in forecasting ICP crises,^[8,9] and ML regression models using computed tomography (CT)-derived features such as ventricular size and optic nerve sheath diameter achieved favorable accuracy with XGBoost.^[10]

At present, computer vision (CV) models have been studied in neurosurgical fields, such as image classification, tumor classification, neuroanatomic segmentation, and surgical posture analysis.^[11-13] Danilov et al., employed the CV model to assess the accuracy of automatically recognizing a neurosurgeon’s optimal posture and hand location.^[11] In contrast, DL-based image models were utilized to classify types of pineal tumors using magnetic resonance imaging from prior studies.^[11] For ICP prediction, there was a lack of evidence regarding the ability to predict ICP value from medical images in the literature review. Therefore, the present study aims to compare the predictive performance of several DL architectures for ICP estimation in patients with HCP.

MATERIALS AND METHODS

The present study’s procedure is demonstrated in Figure 1. Patients with HCP who underwent EVD were enrolled in the hospital database between January 2019 to December 2024. Therefore, patients who were excluded were those without pre-operative CT scans, patients with craniectomy, infants with fontanelles that did not close, and patients with an absent intraoperative ICP value from the operative record. Hence, clinical characteristics and a cranial CT scan were obtained from the remaining individuals. In addition, intraoperative ICP values were collected from electronic medical records, and cranial CT scans of HCP patients were collected.

Close

Figure 1:

Workflow of the present study. MAE: Mean absolute error, MSE: Mean squared error, RMSE: Root mean squared error, R²: Coefficient of determination.

Export to PPT

Cranial CT scans for investigation

Images of cranial CT scans were retrieved from the institutional picture archiving and communication system and de-identified. Axial levels were selected by protocol and verified by a neurosurgeon (10 years’ experience). We selected four clinically meaningful axial levels (the largest lateral ventricular span, the foramen of Monro, the basal cisterns, and the fourth ventricle) that capture ventricular size and cisternal patency relevant to ICP.^[10] Therefore, four axial pictures of cranial CT scans were gathered as follows: The lateral ventricle’s greatest distance [Figure 2a], the level of the foramen of Monro [Figure 2b], the level of the basal cistern [Figure 2c], and the level of the fourth ventricle [Figure 2d]. Total images were subsequently randomly split with a 7:2:1 ratio into a train, validation, and test dataset.

Close

Figure 2:

Cranial computed tomography scans with an axial view for investigation. (a) The lateral ventricle’s greatest distance level (white arrow) (b) Foramen of Monro level (white arrow), (c) basal cistern level (white arrow), and (d) fourth ventricle level (white arrow).

Export to PPT

RESULTS

The present work included 280 individuals with HCP who received EVD between January 2019 and December 2024. Therefore, 16 patients without preoperative CT scans, 13 patients with decompressive craniectomy, eight neonates with fontanelles that did not close, and one patient with an absent intraoperative ICP value from the operating record were eliminated. Consequently, clinical features and cranial CT scan images were obtained for 242 patients, as illustrated in Table 1.

The present study was characterized by a female predominance, with an average age of 25.90 years (±20.58). Common underlying diseases were hypertension, diabetes mellitus, and dyslipidemia, whereas motor weakness was the most common symptom. HCP was frequently caused by intracerebral hemorrhage, subarachnoid hemorrhage, and cranial tumors. In addition, the average ICP was 22.86 cmH₂O (±7.39).

Training model and validation

Several DL architectures were utilized to train models; training and validation losses are shown in Figure 3a-d. CNN train models’ training and validation losses progressively declined, but LeNet and ResNet’s validation losses rose over epochs.

Close

Figure 3:

Training loss by deep learning models (a) convolutional neural network, (b) dense convolutional network, (c) LeNet, and (d) residual neural network.

Export to PPT

Model predictive performance with the test dataset

Predictions and actual ICP values were plotted according to CNN, DenseNet, LeNet, and ResNet architectures using a test dataset, as shown in Figure 4a-d. CNN’s scatter plot shows that it is more evenly distributed along a 45° line than other architectures. In addition, MAE, MSE, RMSE, and R² were calculated and compared among DL architectures in Table 2. When compared to other models, the CNN model had the lowest error parameters, and its prediction had the highest R² value. In detail, the CNN model demonstrated MAE, MSE, and RMSE of 4.96 cmH₂O, 38.76 cmH₂O, and 6.22 cmH₂O, respectively, and it could explain approximately 18% of the variance in ICP value. For implementation, the CNN model was subsequently deployed as a web application through https://cticp123.streamlit.app/, as shown in Figure 5.

Close

Figure 4:

Scatter plot between prediction and actual intracranial pressure value by deep learning models (a) convolutional neural network, (b) dense convolutional network, (c) LeNet, and (d) residual neural network.

Export to PPT

Close

Figure 5:

Screenshot of web application for intracranial pressure prediction. ICP: Intracranial pressure. cmH₂O: Centimeters of water

Export to PPT

Table 2: Predictive performances of deep learning architectures for predicting intracranial pressure.

Architectures	MAE (cmH2O)	MSE (cmH2O2)	RMSE (cmH2O)	R2
Convolutional neural network	4.96	38.76	6.22	0.18
DenseNet	6.58	69.73	8.35	−0.45
LeNet	6.90	82.62	9.09	−0.72
ResNet	6.61	70.82	8.41	−0.48

CNN: Convolutional neural network, DenseNet: Dense convolutional network, MAE: Mean absolute error, MSE: Mean squared error, ResNet: Residual neural network, RMSE: Root mean square error, R2: R-squared

DISCUSSION

ICP prediction has been explored to enhance neurological critical care, particularly in resource-limited settings where invasive monitoring imposes economic burdens.^[14,15] Several approaches have used clinical features, neuroanatomical indices, and cranial CT scans. Alali et al., developed clinical prediction rules incorporating Marshall classification, midline shift >5 mm, GCS ≤4, pupillary asymmetry, and abnormal reactivity, achieving an AUC of 0.86.^[16]

DL-based regression based on medical images has been challenged; the present study compared the predictive performance among various DL models for predicting ICP value in HCP patients. As a result, the model of CNN architecture had the best performance with the lowest error and the highest R². Nevertheless, the R² values from all models in the current investigation were low. This could be explained by the distribution of ICP values in our dataset was imbalanced. A large proportion of patients presented with moderate-to-high ICP, while relatively few cases fell within the very high range. This restricted variability may have constrained the regression models’ ability to capture patterns across the full ICP spectrum, effectively biasing predictions toward high values. Therefore, larger, prospective, and multi-center studies are essential to validate the feasibility of CT-based ICP regression and improve generalizability. However, other error parameters were close to prior studies.^[10,17-19] Trakulpanitkit and Tunthanathip used ventricular indices and optic nerve sheath diameter to predict ICP in HCP patients, reporting XGBoost performance with MAE 3.62– 3.89 cmH₂O, RMSE 6.46–6.72 cmH₂O, and R² 0.50–0.51.^[10] Similarly, Fong et al., developed an ML-based algorithm for ICP crisis detection, achieving RMSE 3.56–4.51 mmHg.^[19] In contrast, our CNN model achieved MAE 4.96 cmH₂O, RMSE 6.22 cmH₂O, and R² 0.18. The lower explanatory power may reflect that ventricular index–based methods target specific anatomic predictors, whereas DL models automatically extract features from the whole image, potentially diluting ICP-relevant signals.^[20,21]

An advantage of regression-based ICP prediction is that it yields continuous values, which can be converted to cerebral perfusion pressure (CPP). Since CPP is critical for maintaining cerebral blood flow, optimal CPP–targeted therapy has been advocated, especially in traumatic brain injury, with guidelines recommending 60–70 mmHg to minimize ischemia and edema.^[22-24] Multimodal ICP prediction combining clinical features, CT imaging, and neuroanatomical indices may serve as a screening tool to guide patient-specific management.^[25,26] Moreover, predictive DL models could be integrated into computerized clinical decision support systems (CDSS) through web applications. A systematic review by Souza et al., reported that CDSS improved care processes, clinical outcomes, costs, and patient safety.^[27]

To the best of our knowledge, this is the first study to predict ICP directly from cranial CT scans. Several limitations must be acknowledged. The optimal number of training images for CNNs remains debated; Shahinfar et al. suggested 150–500 per class for classification tasks^[28] but evidence for image regression is limited. Larger multi-center datasets are needed to improve training, optimize models, and allow robust validation.^[28-30] Future external validation on unseen data will also be essential to confirm the generalizability of our findings.

CONCLUSION

DL regression used to cranial CT demonstrated limited explanatory performance for estimating ICP and is not sufficient as a substitute for invasive monitoring. However, these findings provide preliminary feasibility evidence and a challenge for improvement. Future research using larger, balanced, multi-center datasets, and incorporating multimodal data will be necessary to enhance predictive accuracy and clinical applicability.