Automated deep learning pipeline for callosal angle quantification

Abstract

Background

Normal pressure hydrocephalus (NPH) is a potentially treatable neurodegenerative disorder that remains underdiagnosed due to its clinical overlap with other conditions and the labor-intensive nature of manual imaging analyses. Imaging biomarkers, such as the callosal angle (CA), Evans Index (EI), and Disproportionately Enlarged Subarachnoid Space Hydrocephalus (DESH), play a crucial role in NPH diagnosis but are often limited by subjective interpretations. To address these challenges, we developed a fully automated and robust deep learning framework for measuring the CA directly from raw T1 MPRAGE scans.

Methods

Our method integrates two complementary modules. First, a BrainSignsNET model is employed to accurately detect key anatomical landmarks, notably the anterior commissure (AC) and posterior commissure (PC). Preprocessed 3D MRI scans, reoriented to the Right Anterior Superior (RAS) system and resized to standardized cubes while preserving aspect ratios, serve as input for landmark localization. After detecting these landmarks, a coronal slice, perpendicular to the AC-PC line at the PC level, is extracted for subsequent analysis. Second, a UNet-based segmentation network, featuring a pretrained EfficientNetB0 encoder, generates multiclass masks of the lateral ventricles from the coronal slices which then used for calculation of the CA.

Results

Training and internal validation were performed using datasets from the Baltimore Longitudinal Study of Aging (BLSA) and BIOCARD, while external validation utilized 376 clinical MRI scans from Johns Hopkins Bayview Hospital, as well as PENS trial. Our framework achieved high concordance with manual measurements, demonstrating a strong correlation (r = 0.98, p < 0.001) and a mean absolute error (MAE) of 3.26 (SD 1.89) degrees. Moreover, error analysis confirmed that CA measurement performance was independent of patient age, gender, and EI, underscoring the broad applicability of this method.

Conclusions

These results indicate that our fully automated CA measurement framework is a reliable and reproducible alternative to manual methods, outperforms reported interobserver variability in assessing the CA, and offers significant potential to enhance early detection and diagnosis of NPH in both research and clinical settings.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12987-025-00750-w.

Background

Normal pressure hydrocephalus (NPH) is a potentially treatable neurological condition characterized by the triad of cognitive decline, gait disturbances, and urinary incontinence, as initially described by Hakim and Adams [1]. Epidemiological studies reveal a wide range of prevalence estimates for NPH, but there is a consensus that it remains significantly underdiagnosed [2, 3]. This underdiagnosis is particularly concerning given the treatable nature of the disease [4]. One of the major challenges in detecting NPH is its clinical overlap with other neurodegenerative diseases, such as Alzheimer’s Disease (AD), leading to frequent misdiagnoses [5]. Furthermore, NPH is not often considered a primary differential diagnosis in elderly patients, despite its prevalence [6]. This diagnostic gap highlights the need for enhanced tools and methodologies to increase awareness and improve the identification of NPH cases [7].

Imaging biomarkers play a crucial role in the diagnosis of NPH [8–10]. Measurements such as the Evans Index (EI) [11], Callosal Angle (CA) [12], and Disproportionally Enlarged Subarachnoid Spaces (DESH) [13], provide objective evidence that supports clinical evaluation. Despite their diagnostic value, these imaging methods are often underutilized in clinical practice, primarily due to the perceived rarity of the condition and the time-intensive nature of manual image analysis. As a result, many potential cases may go undetected, emphasizing the necessity for automated, reliable, and efficient imaging tools to assist in the diagnosis of NPH [14]. Implementing such tools can facilitate the identification of new patients and enable the retrospective analysis of previous scans, ensuring that individuals receive appropriate treatment [15].

Recent advancements in artificial intelligence (AI)-based algorithms, coupled with the growing availability of medical imaging datasets and improved computational power, have paved the way for automating complex medical imaging tasks [15]. In this study, we aim to develop a robust AI-based system to automatically measure the CA from raw 1.5T and 3T MPRAGE MRI scans [15]. By automating this critical measurement, our method seeks to improve the sensitivity and specificity of NPH diagnosis, reduce diagnostic delays, and assist clinicians in identifying potential cases, ultimately enhancing the care provided to patients with this treatable condition.

Methods

Data collection

T1-weighted MRI data for training and internal validation were obtained from the Baltimore Longitudinal Study of Aging (BLSA) [16] and BIOCARD cohorts, comprising 1,893 scans acquired on GE and Philips systems across 1.5T SPGR and 3T MPRAGE protocols. External validation included 216 clinically acquired Siemens 3T MPRAGE scans from the Johns Hopkins Hydrocephalus Clinic as well as 160 pre- and post-shunt scans from the PENS trial [17], which contributed a broad range of GE, Philips, and Siemens 3T MPRAGE acquisitions across multiple institutions and imaging protocols. An additional set of 40 S 1.5T T1 spin-echo scans from Johns Hopkins was used exclusively for sensitivity analysis. Complete scanner characteristics, including manufacturer, field strength, sequence, acquisition parameters, voxel resolution, number of scans, and cohort usage, are summarized in Supplementary Table 1. All images were anonymized and converted to NIfTI format using dcm2niix.

Data labelling

The 3D MRIs were anonymized and then annotated for the localization of the anterior commissure (AC) and posterior commissure (PC) coordinates using RilContour software [18]. To accurately measure the CA, the coronal slice represented by the plane perpendicular to the line connecting AC to PC at the level of PC was extracted and saved as PNG images. These coronal slices were then segmented and annotated using Label Studio software to outline the right and left lateral ventricles. The CA was measured according to the standard guidelines for CA calculation. All manual annotations and measurements were thoroughly verified by an experienced neurologist, thereby ensuring the highest standards of data accuracy and reliability.

Quality control

Because the goal of this study was to develop a robust tool applicable to routine clinical imaging, we did not exclude scans on the basis of motion artifacts, noise, contrast variability, or other common acquisition imperfections. Instead, quality control focused solely on criteria essential for accurate preprocessing and valid CA measurement. Scans were included only if they possessed complete and valid structural metadata, including voxel dimensions, spacing information, and orientation, and were acquired as T1-weighted structural images. In addition, scans with incomplete field of view affecting the visualization of the important structures were excluded, and for CA-specific performance evaluation (external validation), images with an EI < 0.30 were removed because CA is only clinically interpretable in the presence of ventriculomegaly. Although the models are optimized for T1-weighted MPRAGE sequences, they were trained and internally validated on a heterogeneous mixture of MPRAGE and SPGR scans, and performance was further evaluated on both MPRAGE and non-MPRAGE T1 acquisitions. No scans were excluded due to motion, intensity inhomogeneity, or noise, reflecting the real-world imaging conditions in which the tool is intended to operate.

Landmark detection model

To automate the detection of the AC and PC landmarks, which are essential for accurate CA measurement, we adapted the BrainSignsNET [19] architecture and modified its output layer to predict only the two required landmarks while preserving the pretrained encoder weights. Prior to model inference, each MRI scan was converted to RAS orientation using the header-defined affine, a rigid transformation that permutes axes without altering voxel geometry or affecting oblique acquisitions. The native voxel spacing and physical dimensions of each scan were then extracted, and the volume was resampled using trilinear interpolation such that the largest physical dimension corresponded to 128 voxels. This preprocessing step produces an isotropically scaled intermediate volume with preserved anatomical proportions. After resampling, each scan is center-padded with zeros to obtain a fixed 128 × 128 × 128 input for the landmark-detection network. Intensity normalization is then applied independently to each scan, using per-scan z-score normalization computed across the entire resampled volume. No global dataset-level statistics or brain-masked intensity constraints are used, a design choice aimed at maintaining robustness across heterogeneous scanners and acquisition protocols. All predicted landmark coordinates are subsequently transformed back into the native physical space using the inverse of the resampling and padding operations.

The modified model was trained and fine-tuned on 80% of manually annotated scans from the BLSA and BIOCARD cohorts, following procedures consistent with the original BrainSignsNET training paradigm. Model selection was based on performance on the remaining 20% of unseen BLSA/BIOCARD data. The final model was then validated on the external validation cohorts. All evaluation metrics were computed by comparing predictions transformed back into native space with landmarks annotated directly on the raw MRI volumes [19].

Ventricle segmentation model

For the segmentation task, we employed a UNet architecture [20] integrating various pretrained backbone architectures. For the final model, the encoder was based on a pretrained EfficientNetB0, while the decoder was initialized with random weights. Preprocessed coronal MRI slices, resized to 224 × 224 pixels and standardized, served as inputs, and multiclass masks of lateral ventricles were the output of the model. The final convolution layer produces three-channel logits per pixel (background, left ventricle, right ventricle) that are passed through a softmax to obtain class probabilities and then converted into the discrete multiclass segmentation mask.

A dual loss function was used to optimize performance: combining the Dice Score Coefficient (DSC) with the Hausdorff Distance (HD). The HD component was specifically incorporated due to its enhanced sensitivity to the segmentation mask borders, which are of critical importance in this project.

To further improve model generalizability, data augmentation techniques, including random rotation, flipping, blurring, noises, skull-stripping, and intensity domain transfer, were applied. The model was initially trained on 80% of the images extracted from the BLSA and BIOCARD cohorts, with the remaining 20% reserved for internal validation. The dataset was split into training and internal validation subsets at the participant level, ensuring that all scans from a given individual were assigned entirely to either training or validation to prevent data leakage. The split was randomly stratified across the BLSA, BIOCARD (1.5T), and BIOCARD (3T) datasets to maintain comparable scan distributions in both subsets. Finally, the best performing model was validated using external validation cohorts.

Callosal angle calculation

To compute the CA the following steps were applied, as illustrated in Fig. 1:

Fig. 1.

Overview of the automated processing pipeline for callosal angle measurement from raw MRI data: 3D MRI scans are first reoriented to the RAS coordinate system, resized to 128 × 128 × 128 cubes (preserving aspect ratio), and transformed using standard scaling (Step1). The anterior and posterior commissures (AC and PC) are then automatically detected using the trained BrainSignsNET model (Step 2), and a coronal slice, perpendicular to the AC-PC line at the level of the PC, is extracted (Step 3). Segmentation masks of the lateral ventricles are generated using UNet model (Step 4), and their marginal contours are delineated. Subsequently, an SVM algorithm computes a separating line with maximum margin between the two ventricles (Step 5a), after which a search algorithm projects perpendicular rays from this line to identify the initial collision points that represent the medial walls (Step 5b). These contact points, satisfying a predefined threshold, serve as base points from which a line representing the medial wall is derived (Step 6). Finally, vectors from these medial wall lines are used to calculate the callosal angle Step 7

1. MRI Preprocessing: 3D MRI scans were reoriented to Right Anterior Superior (RAS), resized to 128,128,128 cubes preserving the aspect ratio of the original scans and transformed using standard scaler.

2. Landmark Detection: The AC and PC were detected from pre-processed 3D MRI scans using the trained BrainSignsNET model.

3. Coronal Slice Extraction: The coronal slice represented by the plane perpendicular to the line connecting AC to PC at the level of PC were extracted.

4. Segmentation and Contouring: Segmentation masks of the lateral ventricles were extracted, and marginal contours were identified using a contouring algorithm.

5. Identification of the medial walls of the lateral ventricles: a: Using Support vector Machine (SVM) algorithm a line separating 2 ventricles with maximum margin were calculated then b: using a specific search algorithm which sends perpendicular lines from the separating line (rays) and calculate first collision point on each ventricle which represent the medial wall of lateral ventricles.

6. Identification of linear representation of medial walls of the lateral ventricles: Points of contact (closest point of each ventricle to the other) meeting a defined threshold between the two ventricles were identified and marked as the base points for each ventricle. Then, using a search method a line starting from each base point that crosses most points of the medial wall of each side were identified.

7. Angle Calculation: Vectors derived from the lines representing medial wall of lateral ventricles were used to calculate the angle between them, representing the CA.

Statistical analysis and evaluation

Numerical demographic data were summarized as mean and standard deviation, while categorical demographic data were reported as counts and percentages.

The performance of the landmark detection model was quantified by calculating the three-dimensional Euclidean Distance between the predicted and manually annotated landmarks. In addition, we computed the Percentage of Correct Key Points (PCK@3 mm), whereby predictions with an error less than 3 mm were classified as correct. Predictions that exceeded this error threshold were deemed incorrect, and this metric was subsequently used to report the model’s overall accuracy in predicting anatomical landmarks. To assess the pitch correction, defined as adjusting the 3D MRI scan such that the AC-PC line is aligned parallel to the horizontal plane in both sagittal and axial axes (ensuring the coronal plane is perpendicular to the AC-PC line), we calculated the angular difference between two lines: one formed by the manually annotated AC and PC points, and the other determined by our predictive method. This angular difference served as the error metric for the pitch correction. The method’s performance was quantified by computing the Mean Absolute Error (MAE). In addition, a Bland-Altman analysis was conducted to visually assess the agreement between manual annotations and predicted alignments. The results of this analysis are reported by both the bias and the 95% limits of agreement.

We evaluated the performance of the segmentation model using the Mean Dice Similarity Coefficient (MDSC) and the Mean Hausdorff Distance (MHD). These metrics quantified the overlap and boundary agreement between the predicted () and ground truth ( segmentations, respectively. MDSC and MHD were calculated using the following formulas:

The final CA measurement was assessed using MAE and Pearson correlation analysis to compare predicted and ground truth values. A threshold of 90° was employed to classify patients into normal or abnormal categories based on both predicted and annotated CA measurements.

To assess the effect of localization and orientation errors on callosal-angle (CA) measurements, we introduced controlled perturbations of the PC landmark and plane orientation. Specifically, we shifted the PC position in 1 mm increments up to ± 10 mm along the anterior-posterior axis and rotated the plane normal in 1° increments up to ± 10° about both the sagittal and axial axes. After each perturbation, the plane was re‐extracted, segmented, and the CA re‐measured. By directly comparing each perturbed measurement to the original, unperturbed CA, we quantified how incremental spatial or angular deviations propagate into CA measurement bias.

Statistical significance was defined as two-tailed p < 0.05.

Training and validation setup

All processing and model training were conducted on a high-performance workstation equipped with dual Nvidia RTX 4090 GPUs, 128 GB RAM, and an AMD Threadripper Pro WRX80 CPU. Training was performed for 100 epochs or until convergence, using the AdamW optimizer with a weight decay of 0.02. For the BrainSignsNet-based landmark detection model, each pretraining epoch required approximately 1123 s, followed by 108 s per training epoch and 18 s per internal validation epoch. All experiments were executed on a compute node utilizing 48 CPU cores and one of two GPU, with a batch size of 6 MRI scans. The model converged after 94 epochs. For the segmentation model, each epoch required approximately 15 s for training and 3 s for internal validation under a similar computational configuration, using a batch size of 16 scans, this model converged after 41 epochs.

Results

Demographics

MRI scans analyzed in this study originated from four cohorts summarized in Table 1. Overall, the study included a total of 1,215 individuals with 2,283 MRI scans. The BLSA cohort (n = 516) had a mean age of 69.31 years (SD = 15.55). The BIOCARD cohort (n = 395) contributed the largest number of scans (1,391 scans) and had an average age of 64.22 years (SD = 11.02), also utilized for training and internal validation. The Johns Hopkins Clinic cohort (n = 216), with a mean age of 71.99 years (SD = 8.09) and PENS [17] (n = 216), with a mean age of 75.06 years (SD = 5.58), provided external validation scans with a Mean EI measurement of 0.37 (SD = 0.05) and 0.39 (SD = 0.03) respectively.

Table 1.

Summarizes the demographic characteristics of participants included in the training and validation datasets

Cohort	Individuals	Number of Scans	Age	Sex	Evans Index	Utilization
BLSA	516	516	69.31 (15.55)	F: 280 (54.4%) M: 236 (45.7%)	-	training and internal validation
BIOCARD	395	1391	64.22 (11.02)	F: 234 (59.24%) M: 161 (40.75%)	-	training and internal validation
Johns Hopkins Clinic	216	216	71.99 (8.09)	F: 99 (45.8%) M: 117 (54.1%)	0.37 (0.05)	External validation
PENS	88	160	75.06 (5.58)	F: 46(49.4%) M: 47(50.6%)	0.39 (0.03)	External validation

In this context, “Age” refers to the participants’ age at the time of MRI acquisition. The table also details gender distribution, the number of scans per dataset, the total number of participants, and the mean Evans index with its standard deviation (mean ± SD)

Landmark detection and MRI pitch correction

The final evaluation of the trained BrainSignsNET model for detecting the 3D coordinates of anatomical landmarks demonstrated high accuracy and robustness. Specifically, internal validation yielded a median error (MED) of 1.8 mm for the AC and 1.57 mm for the PC, External validation showed similar performance, with MEDs of 1.86 mm for AC and 1.97 mm for PC. Moreover, a vast majority of predicted landmarks were within 3 mm of the corresponding annotated landmark, achieving a PCK@3 mm value of 96.25% across all evaluations. These results show significantly better performance compared to the original performance of the pretrained baseline BrainSignsNet in the detection of PC (MEDs of 2.0.9 mm) and AC (2.65). For context, inter-observer variability in manual AC/PC identification was 1.69 ± 1.26 mm for AC and 1.92 ± 1.06 mm for PC (combined mean 1.81 mm) on the external validation dataset.

The performance error of the pitch correction is visualized in Fig. 2a. These results indicate an overall pitch correction characterized by 95% limits of agreement ranging from − 4.33° to + 3.76°, with a bias of -0.28°. Additionally, the Mean Absolute Error (MAE) was 1.81° with a standard deviation of 1.64°. The pitch angle using predicted landmarks had a significant correlation with the manually annotated pitch angle (r = 0.9, p-value < 0.001) Fig. 2b.

Fig. 2.

Overview of the performance evaluation of the automated framework. Panels (a)–(b) and (c)–(d) illustrate the performance of AC-PC angle (pitch) and callosal angle estimation, respectively. Panel (a) shows the Bland-Altman plot assessing the performance error of the pitch correction procedure relative to manual annotations, while panel (b) illustrates the correlation between the pitch angles obtained from the predicted landmarks and those from manual annotations. Panel (c) presents the Bland-Altman analysis comparing the overall automated callosal angle (CA) measurements to the manual CA measurements, and panel (d) displays the correlation between the predicted and manually measured CA values. Additionally, panel (d) indicates in red the CA measurements that would be incorrectly classified as narrow versus wide based on a 90° threshold

Segmentation models

Among the evaluated segmentation architectures, the UNet model equipped with a pretrained EfficientNetB0 encoder consistently demonstrated the highest performance. On the internal validation dataset, it achieved an overall Dice score of 92.49% and a mean Hausdorff distance (MHD) of 1.29 mm, with right and left ventricular Dice scores of 92.31% and 92.68%, respectively. On the external validation dataset, the model maintained strong generalizability, achieving an overall Dice score of 94.91% and an MHD of 1.88 mm. The right and left ventricles achieved Dice scores of 94.90% and 94.93%, with corresponding MHDs of 1.90 mm and 1.86 mm. Comprehensive performance comparisons across all candidate models are provided in Supplementary Table 1. For reference, manual inter-observer variability in ventricular segmentation showed Dice agreements of 94.55% ± 1.85 for the left ventricle and 93.92% ± 2.10 for the right ventricle, indicating that the model’s performance closely matches expert-level reproducibility.

Callosal angle measurement performance

As demonstrated in Fig. 2c, overall performance evaluation of the proposed algorithm in measurement of CA from 3D raw MRI showed a MAE of 2.95 (SD = 1.58) degrees between the predicted CA and the manually measured CA. The Bland-Altman analysis demonstrated that the overall automated CA measurement yielded 95% limits of agreement ranging from − 7.14° to + 8.32°, with a bias of 0.59°. The analysis showed a significant correlation between predicted and manually annotated CA with a correlation coefficient of 0.98 and P-value < 0.001 (Fig. 2d). Moreover, considering a threshold of 90° for classifying narrow versus wide CA, out of 376 scans in the external validation dataset, only 11 scans were classified incorrectly.

To further illustrate the robustness of our pipeline, we conducted a small sensitivity analysis by applying the method to 40 non-MPRAGE scans (Supplementary Fig. 2), recognizing that these sequences are not intended for clinical CA assessment. The predicted CA values were compared to the manual measurements obtained from the corresponding MPRAGE scans. This comparison yielded a mean absolute error (MAE) of 3.3° ± 1.61°. A subsequent Bland-Altman analysis indicated that the 95% limits of agreement ranged from − 9.93° to + 10.56°, with a bias of 0.31° (Supplementary Fig. 2a). Moreover, correlation analysis demonstrated a highly significant relationship between the predicted and manually annotated CA values, with a correlation coefficient of 0.98 (p < 0.001; Supplementary Fig. 2b).

Error and bias analysis

Figure 3 illustrates that neither the pitch correction error nor the CA prediction error exhibited statistically significant correlations with age or the EI. Additionally, no significant differences were found between genders for either error metric (all p-values > 0.05).

Fig. 3.

Evaluation of the error differences in the automated framework stratified by clinical and demographic parameters. Panels (a) and (b) are scatter plots showing the correlation between pitch correction errors and the Evans Index and age, respectively, while panel (c) is a box plot depicting the differences in pitch correction errors by gender (t-statistic = -0.02, p-value = 0.98). Similarly, panels (d) and (e) are scatter plots illustrating the relationship between callosal angle measurement errors and the Evans Index and age, respectively, with panel (f) presenting a box plot of callosal angle measurement errors stratified by gender (t-statistic = 0.27, p-value = 0.78)

We also evaluated whether scanner variability introduced systematic bias into model performance. Despite substantial heterogeneity in vendor, field strength, and pulse sequence across datasets, no significant differences were observed in segmentation accuracy or landmark-detection error. Performance was comparable between BLSA and BIOCARD scans (Dice: 0.92 vs. 0.90; landmark error: 1.73 mm vs. 1.80 mm) and between Philips 3T MPRAGE and GE 1.5T SPGR acquisitions (Dice: 0.91 vs. 0.89; landmark error: 1.81 mm vs. 1.80 mm), with all p-values > 0.05. Importantly, the model also generalized effectively to an entirely unseen, clinically acquired Siemens 3T dataset, where segmentation and landmark-localization errors were comparable or lower than those observed in the research cohorts. In the external validation, we additionally found no significant differences in landmark-detection accuracy, pitch-correction error, or CA estimation across scanner manufacturer, acquisition sites, or pre- vs. post-shunt status (all p-values > 0.05). The segmentation model demonstrated higher DICE score in the PENS trial relative to the Johns Hopkins clinical cohort, which is expected given the larger ventricles (EI) in the PENS population (p < 0.01). Together, these results indicate minimal scanner-related bias and robust cross-scanner generalizability of the proposed pipeline.

In the sagittal-offset analysis (Fig. 4a), introducing a 5° pitch error increased the CA measurement error to approximately 6.13 ± 4.45° (clockwise) and 4.77 ± 3.62° (counterclockwise). When the pitch error was doubled to 10°, the measurement error rose further to 9.21 ± 6.64° (clockwise) and 7.63 ± 6.20° (counterclockwise). In the axial‐offset analysis (Fig. 4b), a 5° rotation increased the callosal‐angle error to 3.26 ± 2.67° (clockwise) and 3.91 ± 3.10° (counterclockwise). A 10° axial error showed 4.76 ± 4.20° (clockwise) and 4.74 ± 3.94° (counterclockwise). Finally, shifting the PC landmark posteriorly by 5 mm increased the CA error to 10.60 ± 6.99°, and a 5 mm anterior shift yielded 12.95 ± 8.26° (Fig. 4c). At the extremes (± 10 mm), posterior displacement produced 16.58 ± 8.91° of error, while a 10 mm anterior shift produced 25.63 ± 14.45°. In all three perturbation experiments, each nonzero deviation was significantly different from baseline (paired t‐tests, p ≪ 0.001) (Supplementary Fig. 1).

Fig. 4.

Boxplots of absolute callosal-angle (CA) measurement error as a function of perturbation magnitude for three sources of potential bias: sagittal‐plane angle offsets (a), axial‐plane angle offsets (b), and posterior commissure (PC) landmark shifts (c). For each subplot, the x‐axis indicates the magnitude of the perturbation (in voxels (equals to 1 mm) for PC shifts; in degrees for sagittal and axial rotations), and the y‐axis shows the resulting absolute CA error (°) compared to the unperturbed baseline. Asterisks above each box denote the level of statistical significance (* p < 0.05; ** p < 0.01; *** p < 0.001) from paired t‐tests against the zero‐perturbation condition

Discussion

In this study we introduce a fully automated framework for the measurement of the CA from raw MRI data according to the CA measurement method described by Ishii et al.[21], offering an objective and reproducible approach to quantifying this clinically significant index. The framework synergistically integrates two deep learning models designed to perform complementary tasks. The first model detects key anatomical landmarks in three-dimensional MRI data, specifically, the anterior and posterior commissures, while the second model executes two-dimensional ventricular segmentation to delineate accurate ventricular contours. By combining these techniques, the framework establishes a robust methodology for CA measurement, significantly reducing reliance on manual, error-prone procedures. Extensive evaluations on both internal and external validation datasets, comprising research-grade as well as clinically acquired MRI scans, demonstrate the method’s high accuracy and generalizability. The framework not only exhibits robust performance on previously unseen clinical T1 MPRAGE scans as well as both pre- and post-shunt surgery scans from the multicenter PENS cohort for CA calculation, but also maintains acceptable performance on non-MPRAGE clinical scans, which are commonly used in clinical settings for brain evaluation. Furthermore, bias analysis indicates that age, sex, the severity of hydrocephalus (as quantified by the EI), and scanner characteristics are not associated with model performance, thereby ensuring its applicability especially to different scanners and patients with a wide range of ventricular volumes (EI > 0.3).

Previous studies have proposed methods to evaluate MRI metrics related to NPH, including the EI [22, 23], CA [15], and DESH [24], as well as approaches for direct NPH diagnosis from MRI scans. Many of these methodologies rely on either manual procedures or automated segmentation tools, such as FreeSurfer [25] for three-dimensional brain segmentation, which often lack generalizability when applied to clinically acquired MRI scans, particularly those deviating from standard protocols or involving atypical cases like severe ventricular enlargement or hydrocephalus. In these contexts, segmentation errors become more pronounced and critical anatomical landmarks, such as the anterior and posterior commissures, are not properly identified, leading to discrepancies in indices such as the CA. Furthermore, studies have demonstrated that even minor deviations in pitch correction, errors of less than 10° relative to the AC-PC plane, can significantly alter CA measurements [26]. In our study, we similarly confirmed that small misalignments or landmark-localization errors produce significant CA bias. For instance, modest sagittal‐plane rotation or axial tilt each introduced several degrees of angular error, and shifting the PC landmark (coronal slice location) by only a few millimeters resulted in substantial deviations in CA measurement. These findings underscore that precise, consistent localization and alignment are critical to obtaining valid callosal‐angle measurements. A similar concept applies to other MRI-derived measurements, such as the EI. The EI is calculated on the axial plane that captures the maximal width of the anterior horns of the lateral ventricles, a plane that must be perpendicular to the AC-PC line. Consequently, accurate detection of these anatomical landmarks is imperative for the reliable computation of various MRI-driven indices. Our proposed framework directly addresses this challenge by automatically detecting the AC and PC with high precision, aligning the scan according to the true AC-PC orientation, and extracting the coronal slice perpendicular to this line at the level of the PC. As demonstrated in our pitch-correction analysis (Fig. 2a-b), this automated alignment closely matches manual alignment and ensures that the CA is consistently measured in its anatomically correct plane, thereby overcoming the limitations of previous methods.

In this study, we developed a robust and externally validated deep learning framework for CA measurement that leverages a combination of landmark detection and segmentation algorithms to extract precise anatomical landmarks and delineate ventricular structures. This integrated approach adheres closely to standardized methodologies, ensuring a transparent CA measurement process that facilitates the visualization and identification of potential errors, thereby enhancing clinical utility. Our framework demonstrated accuracy comparable to manual measurements and broad applicability across both research and clinical datasets, including MRI scans from patients with markedly enlarged ventricles as our analysis showed no significant correlation between the error related to pitch correction and CA measurements and severity of hydrocephalus evaluated by EI. Moreover, all models performed well in both internal and external validation datasets, ensuring the robustness and generalizability of the method.

Furthermore, our method exhibits superior performance relative to previous automated CA-measurement approaches. Prior to CA estimation, both landmark detection and ventricular segmentation demonstrated accuracy comparable to human raters. These values define the practical precision limits of the upstream steps that directly influence CA measurement. Despite this inherent human variability, our framework achieved markedly lower CA variability than previously reported methods: while a prior study reported 95% limits of agreement of − 29.99° to + 21.98° and radiologist interobserver variability of − 14.95° to + 7.39°, our approach demonstrated substantially tighter limits of agreement from − 7.14° to + 8.32°. These findings underscore the robustness and clinical reliability of our fully automated pipeline for CA assessment in NPH-related MRI [15]. This performance was consistently observed across a large multicenter cohort as well as clinically acquired pre- and post-shunt scans, demonstrating the robustness of the proposed framework in real-world clinical settings. These findings support the utility of the method both as a diagnostic tool following the radiological detection of ventriculomegaly (EI > 0.3) and as an objective, reproducible instrument for evaluating patients after shunt placement.

Despite the promising results, this study has several limitations. Full scanner-wise harmonization and intensity-distribution analyses could not be performed for the BLSA dataset because raw DICOM files were unavailable without a lengthy request process, limiting formal cross-scanner bias assessment to the BIOCARD cohort, the Johns Hopkins clinical dataset, and the multicenter PENS trial. Although no systematic scanner-, sequence-, or field-strength-dependent errors were observed across these datasets, broader evaluation across additional vendors and acquisition protocols remains necessary before clinical deployment. Moreover, several analyses, such as scanner-wise intensity histograms before and after normalization, and KL-divergence quantification, were not included and should be incorporated to ensure consistent performance prior to clinical implementation. The framework, while highly accurate for AC-PC detection, pitch correction, ventricular segmentation, and CA measurement, is not a standalone diagnostic tool, and its integration with clinical assessments and additional imaging biomarkers will require validation in prospective studies. Furthermore, the pipeline was trained and validated exclusively on T1-weighted MRI; non-T1 modalities were not evaluated, and low-resolution T1 spin-echo scans were included only as a limited sensitivity analysis. Future work should integrate additional imaging biomarkers, multimodal inputs, and comprehensive failure-mode analyses across larger and more heterogeneous clinical datasets. Finally, because the CA is clinically meaningful only in the presence of ventriculomegaly, CA performance in this study was evaluated exclusively on scans with EI > 0.3; thus, the method is not designed for, nor should it be applied to, individuals without ventricular enlargement, consistent with the intended diagnostic use of this biomarker.

Conclusion

This study presents a robust and fully automated AI-based framework for CA measurement directly from raw MRI, integrating accurate AC-PC landmark detection, reliable pitch correction, and high-quality ventricular segmentation. By training and validating the pipeline across a diverse collection of scanner vendors, field strengths, acquisition protocols, and both research-grade and clinically acquired datasets, including multicenter pre- and post-shunt imaging, the method demonstrates strong generalizability and vendor-neutral performance. These characteristics make the system suited for real-world deployment, where imaging heterogeneity is the norm. Importantly, the automated framework substantially reduces user-dependent variability, enabling consistent callosal-angle measurement at scale and supporting earlier and more reliable identification of patients with NPH. Together, these findings represent a meaningful advancement toward clinically integrated, AI-driven tools that can enhance diagnostic accuracy, streamline radiological workflows, and ultimately improve patient care in NPH and related neurological disorders.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

S.S.B. developed the code, performed the analyses, proposed the methodological framework, and wrote the main manuscript text. A.M. conceptualized the study, supervised the project, and contributed to the interpretation of results and manuscript revisions. M.B. refined the methodology, assisted with analysis, and contributed to interpretation. M.S.A., M.I.M., M.G.L. provided data resources, scientific expertise, and critical input on study design and interpretation. C.V., L.A., and A.Mo. contributed to data annotation and preparation. All authors reviewed and approved the final version of the manuscript.

Funding

This work was supported by the National Institute on Aging (Grant Numbers U19-AG033655 and P30-AG066507), and the National Institute of Neurological Disorders and Stroke (U01-NS12276).

Data availability

The complete codebase for Automated Callosal Angle Measurement, including all scripts and implementations of the BrainSignsNet, 2D Ventricular Segmentation, and Callosal Angle Measurement code, is publicly available at https://github.com/SiavashShirzad/AutomatedCallosalAngle. The pretrained model weights used in this study will be provided upon reasonable request from the corresponding author and are planned to be released publicly in a future update.

Declarations

Ethical approval

All procedures involving human participants were conducted in accordance with the ethical standards of the institutional and/or national research committees and with the 1964 Helsinki Declaration and its later amendments. The Baltimore Longitudinal Study of Aging (BLSA) and BIOCARD studies were approved by the Institutional Review Board (IRB) of the National Institute on Aging (NIA), and all participants provided written informed consent at the time of enrollment. The use of clinically acquired MRI scans from the Johns Hopkins Hydrocephalus Clinic was approved by the Johns Hopkins Medicine Institutional Review Board with a waiver of informed consent due to the retrospective and fully de-identified nature of the data. The PENS trial data were obtained under IRB approval at participating institutions, and all participants provided informed consent consistent with the trial protocol. All MRI scans used in this work were anonymized prior to analysis, and no identifiable information was accessible to the study team. No new data were collected specifically for this study, and all analyses were performed on existing, de-identified datasets.

Competing interests

The authors declare no competing interests.