Normal pressure hydrocephalus (NPH), is a prominent type of dementia that is often reversible via ventricular shunt surgery, with earlier intervention leading to better outcomes (1). A notable discrepancy between the incidence (of patients who had surgical intervention) (2–4) and prevalence rates (5–7) suggests its under-recognition. This is substantiated by an estimate from the Hydrocephalus Association that 80% of patients with NPH remain unrecognized with most frequent misdiagnoses being Alzheimer’s Dementia (AD) or Parkinson’s Disease (PD) (8), which themselves significantly contribute to the global burden of neurological disorders (9). About half of the cases of NPH are estimated to be idiopathic, and the other half secondary to traumatic brain injury (TBI), tumor, meningitis, or infections (10, 11). Adding complexity to its recognition may be the fact that in the secondary NPH group, while tumors or infection linked to the disease can be detected definitively, a TBI-related origin may not be accurately identified due to incomplete medical history. Therefore, the accurate detection and treatment of NPH may inevitably depend on discerning it from irreversible atrophic pathologies like AD, PD, and post-traumatic degeneration.

In terms of structural neurodegeneration, hydrocephalic ventriculomegaly is the key marker of NPH (12). Idiopathic NPH and NPH secondary to TBI, may differ in their etiology, but remarkable improvement in symptoms of patients with both conditions have been demonstrated after ventricular shunting (1, 13). But because atrophy is a chronic effect of TBI which may cause secondary ventriculomegaly, and has been negatively correlated with shunt outcome, accurately discerning it from hydrocephalic ventriculomegaly on structural imaging (14) is paramount for shunt-surgery decision making. The structural degeneration in AD (15) and PD (16) is also characterized by atrophy. Both atrophy and hydrocephalus can give an appearance of ventricular enlargement, which may set a precedent for misdiagnosis if not examined carefully.

In addition to similar structural neurodegeneration, the cognitive and functional deficits such as dementia and gait impairment caused by these diseases often overlap, rendering misdiagnoses distressingly common. The stretching of the corticospinal tract (CST) in the corona radiata which conducts signal to the legs is thought to produce gait disturbance, a manifestation in most NPH cases, while radial shearing force exerted by enlarging ventricles leads to dementia (17). Loss of structural integrity leads to impairment of cognitive and executive function in AD which may affect gait due to divided attention (18). PD which is thought to arise in the substantia nigra and basal ganglia, is characterized by its motor symptoms including gait impairment (19). Its degenerative impact extends well into the cerebral cortex as atrophy leading to cognitive deficits (20). Chronic TBI (cTBI) also leads to cognitive and gait impairment through its degenerative effects (21, 22).

The neurostructural damage resulting from NPH manifests as distinct imaging markers capturable on Computed Tomography (CT) and Magnetic Resonance Imaging (MRI). CT-based markers have been included as a supporting factor in the diagnostic guidelines for NPH and acute TBI. However, they have not gained prominence relative to neurological, clinical, cerebrospinal fluid (CSF) and blood biomarkers, advanced MRI, and functional imaging, especially in AD, PD, and cTBI. Structural imaging markers may offer a noninvasive solution not only for the accurate detection of NPH, but also for its distinction from irreversible neurodegeneration in AD and PD; and distinguishing between atrophy and hydrocephalus in cTBI. There have been a handful of reviews that have elucidated imaging markers of NPH. Pyrgelis et al. (23) provided a review of functional and structural imaging markers for NPH and highlighted a limited number of them which distinguish it from AD and PD, and Yin et al. (24) presented a mini-review of NPH-specific features. But there have been no focused attempts to highlight the structural imaging markers of NPH along with those that can differentiate it from its mimics AD and PD and shed light on the interplay of atrophy and hydrocephalus in cTBI. This review also comments on the adoption of computational techniques for the objective assessment of these markers and highlights areas where further research is needed.

We searched PubMed for the keyword “normal pressure hydrocephalus” in combination with the keywords “Alzheimer” and “parkinson.” We also searched for “hydrocephalus” and “atrophy,” in combination with “traumatic brain injury.” Studies that used only structural neuroimaging modalities of (T1/T2/DTI) MRI and CT, and those that studied humans aged 19 years and above were retained. From the resultant set of articles, we present this review of structural imaging markers for NPH. For AD and PD, we provide a brief background of their burden, pathology, structural degenerative markers that differentiate NPH from them. The review of cTBI, and consequent atrophy and hydrocephalus is structured similarly. Articles where (semi) automatic image processing and machine (deep) learning have been applied to assess these features in MRI and CT modalities are also discussed. The key abbreviations and structural imaging marker definitions are in Tables 1, 2 respectively.

Hakim and Adams, first explained the mechanism of NPH using Pascal’s law. Presently, at least 700,000 older adults in the US are estimated to be afflicted by the condition (40). Excessive buildup of CSF in the ventricles leads to their enlargement and impingement on the surrounding brain tissue. And due to the compliance of the surrounding brain tissue, intracranial pressure (ICP), remains “normal” (12). Recent studies suggest that the pathogenesis of NPH is complex and is potentially related to CSF dynamics, CSF – interstitial fluid exchange, cortical subarachnoid space (SS) morphology, and venous congestion (41). It is clinically characterized by the Hakim triad which is a combination of gait impairment, urinary incontinence, and cognitive impairment.

The American-European diagnostic guidelines classify NPH into possible and probable subgroups based on clinical evaluation for the Hakim triad, an invasive CSF tap-test with opening CSF pressure of 5–18 mm Hg, medical history, and radiographic assessment of lateral ventricle (LV) enlargement (42). Enlarged temporal horns (TH) and periventricular intensity changes unattributable to ischemia or demyelination are also considered. While they do not endorse a classification of patients with definite NPH based on positive shunt response, the Japanese guidelines take that approach, and include evaluation of sulcal tightness at the high-convexity (HC) of the brain and dilation at the sylvian fissures (SF) (43). Age at onset and presence of other comorbidities are also considered in both guidelines (44). Irrespective of different approaches to classification, a vast majority of patients saw favorable outcomes following shunt surgery (45) and early surgical intervention was found to significantly affect functional outcomes (46). And despite reports of complications (47), advances in surgical procedures guided by computer aided neuronavigation, and infection reduction strategies suggest a promising future for complete reversal of symptoms in NPH (48).

Alzheimer’s Dementia (AD) is best characterized by the abnormal presence of extracellular amyloid plaques and intracellular neurofibrillary tangles (NFT) of protein tau which leads to a loss of synapses and cell death (78). In the US, AD is prevalent among 6.5 million people (79). Studies have shown that amyloid beta plaques start accumulating in abnormal clumps called oligomers and fibrils in the early stages which is correlated to downstream accumulation of NFT (80). Activated microglia that are unable to keep up with debris clearance cause chronic inflammation and cell death leads to atrophy. In early stages, the medial temporal lobe involving hippocampus, entorhinal cortex, amygdala, and para-hippocampal cortex seem to be most affected along with global atrophy involving the frontal, temporal, and parietal lobes excluding the sensory-motor cortex and occipital regions. Basal ganglia regions like the caudate, putamen, pallidum, and nucleus accumbens also show atrophy (15). Clinical symptoms mainly include loss of memory, decreased executive function, language deficit, loss of vision, and gait instability (78). Diagnosis is based on physical exam, evaluation of patient history, memory, cognitive, and neurological functions along with radiographic imaging. Therapeutic solutions for AD like cholinesterase inhibitors are commonly used to alleviate symptoms. While there is no complete cure for the disease yet, a drug which was shown to remove amyloid plaque deposits in the brain was approved by the FDA and several others are being investigated (79).

Parkinson’s disease (PD) is primarily diagnosed in a clinical setting through thorough examination of neurological symptoms that have motor manifestations like tremor, impairment of gait, arm-swing, balance, postural stability, and facial expression, and non-motor manifestations like impairment of behavior and cognition. Definitive diagnostic tests using blood, CSF, or imaging biomarkers do not currently exist (87). The manifestation of dementia in PD (PDD) patients (16) makes some patients more susceptible to misdiagnosis. There are reports that indicate that more than 15% of patients with PD may be misdiagnosed (88), with prominent misdiagnoses including AD (16). NPH may also be misdiagnosed as PD due to parkinsonism (89). Early diagnosis of this disease is challenged by the late manifestation of its defining motor symptoms, heterogeneity in clinical presentations, underlying mechanisms of its subtypes, age at onset, rate of progression, and response to treatment. This is also reflected in the heterogeneity of imaging features which correlate with symptomatic presentation (87).

As a direct consequence of the fact that the source of this synucleinopathy is known to primarily affect the substantia nigra (SN) and basal ganglia, and its microscopic magnitude of origin, imaging markers from structural imaging which rely on the scatter of external radiation like CT and MRI are not recommended for assessment of PD. Rather, functional imaging markers that directly correlate with the function at cellular levels by reflecting the uptake of specific radioactive tracers are the most viable option for detection, especially in early stages. In support of this claim is an article from Dalaker et al. (90) who found that global markers like atrophy and white matter hyperintensities were not significantly distinct in early PD as compared to healthy controls. Recently, Bae et al. (91) reviewed the degenerative markers of the SN on advanced imaging techniques like Nigrosome and Neuromelanin Imaging (NMI) using high field MRI techniques, Quantitative Iron Mapping (QMI), and Single Photon Emission Computed Tomography (SPECT). They highlight studies which have shown highly sensitive classification of PD when iron deposition in the SN driven by dopaminergic cell loss is used as a biomarker. In support of high field and neuromelanin sensitive MRI is a review from 2019 by Prange et al. (92) and from 2016 by Pagano et al. (93) Near normal or diffuse cortical atrophy is associated with PD (94), which raises the possibility of no visible markers of degeneration on CT or MRI. We investigate this further by providing the following review of structural imaging markers in PD and those that can distinguish it from NPH on MRI and CT sequences.

Early adopters of semi-automatic algorithms included an effort to extract volumetric ventricular sizes and cortical thickness for distinction of NPH from AD, PD, and normal controls using MRI, and advocated for the ratio of ventricular size and cortical thickness as a better feature to distinguish between definite NPH (n = 5), from the rest, especially as ventricular volumes in NPH and AD may overlap (72). This problem has also been countered by considering the distribution of (normalized) CSF volumes rather than using it as a global measure, as shown by many studies so far. Voxel-based morphometry (VBM) has long been established as a reliable computational tool in MRI studies to assess structural markers of NPH. Yamashita et al. (76) analyzed CSF distribution using VBM on MRI scans and showed significant distinction between NPH (n = 12 with 83% probable NPH and 17% definite NPH) when compared to AD and normal controls in terms of LV/SF volumes, and HC/ML volumes. In a measure that reflects DESH, their analysis also showed that the volume of CSF in the LV and SF areas as compared to that in the HC and ML areas was higher in NPH as compared to AD and normal controls. Subsequently, they showed similar trends in those features on a definite NPH cohort, and when it was compared to PD patients (75).

Ishii et al. (74) developed an automatic volumetric CSF segmentation method on T1-MRI, which provided localization into the SF, HC, and ML sulci, as well as the ventricular spaces. They demonstrated the validation of previously known findings that enlarged ventricles and SF, and tight HC sulci were evident in a group of 15 probable NPH patients as opposed to AD and healthy controls. In a more targeted and automatic effort to segment GM, WM, and CSF spaces on T1-MRI, ventricular and GM volume, and gender were found to distinguish shunt-responsive NPH from AD and normal controls (73). Ellingsen et al. (108) developed a segmentation and labeling software to compartmentalize the ventricular system (into lateral, third, and fourth), referred to as RUDOLPH on MRI at a time when limited or no efforts were made for automatic segmentation of highly deformed ventricles. Yamada et al.’s studies on indices characterizing CSF spaces for the optimal distinction of probable NPH from healthy controls (29), AD, and NPH-and-AD cohorts on T2-MRI (38) were also based on semi-quantitative approaches relying on automatic segmentation of brain tissue. In a recent study of CSF distribution, the normalized WM and CSF volumes derived using VBM, along with the CA were shown to provide improved distinction between probable NPH and AD using MRI sequences when compared to the CA alone (77).

Peterson et al. (55) used an automatic segmentation method on T1-MRI to quantify volume in deep subcortical GM structures and showed reduced volume of the caudate, putamen, thalamus, nucleus accumbens, hippocampus, and pallidum in 14 NPH patients as compared to 8 healthy controls. Efforts apart from volumetry have also been made to extract nuanced MRI-based features. Statistical parametric methods using MD histogram analysis from DTI images have also been shown to distinguish between probable NPH and AD/PD/normal controls with minimal manual intervention (109).

Linear and interpretable measures of ventriculomegaly have also been computationally assessed on MRI. Borzage et al. (7) developed an automated methodology to extract CA from MRI from the Open Access Series of Imaging Studies (OASIS) and Alzheimer’s Disease Neuroimaging Initiative (ADNI) databases. With a high ICC of 0.9 between the automated measure and expert annotations on 281 images and an innovative approach to coronal pitch correction for image standardization, they showed the utility of data-driven applications in objective evaluation of diagnostic standard features. CT based methods for automatic assessment of interpretable measures have been limited. An automated image-processing methodology to extract the EI-x from CT images reported a correlation coefficient of 0.983 between the automated measure and expert annotations among 44 subjects (12 NPH). However, they applied nonlinear registration for image standardization which would render the brain width as a constant and affect the measurement of the EI-x (110).

Deep learning-based methods have shown promise in the objective and automatic assessment of NPH using structural imaging modalities. Again, MRI dominates this field of investigation. To enable the automatic segmentation and parcellation of the ventricular system in patients with severe ventriculomegaly due to hydrocephalus, Shao et al. (111) developed a deep learning model named VParNet trained on MRI scans of NPH patients and healthy controls, which showed high agreement with expert annotations and outperformed state-of-the-art. However, they did not test the classification capacity of their model. Irie et al. (112) developed a 3D convolutional-ladder network, and showed that their model can distinguish between probable NPH (n = 23), AD, and controls with an accuracy, sensitivity, and specificity of about 90% using MRI scans. By showing regions like the periventricular spaces, hippocampus, and THof the LVs being highlighted in the activation maps sensed by the network on some successfully classified scans, they demonstrated agreement with previous neuroanatomical findings in their findings.

Deep learning methods to assess NPH on CT scans have also emerged recently. Considering an EI-x ≥ 0.32 to be an indicator of hydrocephalus, a transfer learning scheme was applied on a large dataset of CT scans to show a classification performance with AUC of 0.93, sensitivity of 93.6%, and specificity of 94.4% in distinguishing hydrocephalus from normal controls (113). Automated segmentation of CSF, subarachnoid, and cerebral spaces on non-contrast CT scans of 27 patients with possible NPH, integrated with indirectly inferred connectome data, was shown to be as effective as the EI-x for prediction with a specificity of 85% and sensitivity of 86% (114). Haber et al. (115) recently developed a convolutional neural network (CNN) which was able to classify patients with definite improvement post-surgery as identified by the 2nd edition of the Japanese guidelines from healthy controls using CT scans setting a promising precedent in the application of deep learning to CT scans in NPH assessment. Further studies are required to study the potential of deep learning in assessing imaging markers that can predict positive and objective shunt response in symptomatic NPH and distinguish it from its mimics.

TBI casts a pervasive shadow, affecting an estimated 27 to 69 million individuals globally each year, with each case carrying significant chronic consequences (116, 117) and pathophysiological connections to various neurodegenerative diseases including NPH, AD, and PD. Its impact can range from concussions and diffuse axonal injury (DAI) which are harder to detect on structural imaging to extra-axial and intracerebral lesions that appear prominently. While acute injury is easily identified on CT and treated surgically, detecting (and differentiating between) its chronic degenerative consequences including atrophy and hydrocephalus which is crucial for optimal management may be challenging.

Subdural hemorrhage (SDH) refers to bleeding that occurs in the subdural space of the brain. And chronic SDHs (cSDH) represent these lesions which may evolve bilaterally in patients with symmetric cranial vaults, but most frequently occur on the side with higher frontal or occipital convexity. They may also be seen at interhemispheric locations in some cases (118). Even though intracranial hypotension and defective coagulation can cause cSDH, mild trauma to the head remains the predominant cause (119). Rupturing bridging veins in the subdural space following trauma may cause hematomas with varying cellular and vascular compositions as they evolve (118). It mostly affects older patients and is diagnosed with CT even though MRI might be more sensitive in isodense cases. Cerebral atrophy has been shown to be a risk factor for cSDH using volumetric analysis of CT scans (120), as well as being a prominent chronic consequence of it (121). Progressive volume loss in cSDH, at rates higher than dementia, have also been reported using volumetric analysis of CT scans (122). While hydrocephalus is also reported to be a risk factor for the development of cSDH following a minor head injury (123), there has been no compelling investigation into its development post cSDH.

Subarachnoid hemorrhage (SAH), which refers to bleeding in the space between the arachnoid and pia mater of the brain, is associated with secondary hydrocephalus (124) and atrophy (125). NPH was shown to develop in 21% of patients within a month following SAH, and improvement was noted in 85% of patients who underwent shunt surgery (126). While SAHs can occur due to nontraumatic causes, it is also estimated to occur in 33–60% of patients with moderate to severe TBI (127). Dilation of the LVs, as assessed with linear measures such as the bicaudate index and dilation rate assessed as change in volume over time on CT (124), atrophy measured indirectly as the modified cella media index (mCMI) and normalized CSF volume on T1-MRI (32), and parenchymal volume loss assessed through volumetric methods (128) are known structural markers of nontraumatic/spontaneous SAH. Hydrocephalus also occurs in traumatic SAH (13, 129, 130) which are usually found in the sulci at the convexities (131). Tian et al. (129) found that it may be correlated with intraventricular bleeding, severity of injury, thickness, and location of the lesion. Overall, we found that there are limited efforts to characterize and differentiate between the structural markers of hydrocephalus and atrophy following traumatic SAH. The prominent features of atrophic and hydrocephalic degradation in cTBI are summarized in Table 5.

Computational techniques like tensor-based morphometry (TBM) have been used on MRI (T1-weighted) scans to study localized volume loss in cTBI. WM structures including but not limited to the CC, subcortical structures like the caudate (middle and posterior) cingulate, thalamus, frontal and temporal neocortices, and the cerebellum, were shown to be affected after at-least 3 months of moderate–severe TBI as opposed to controls using this technology (145). Ventricular volume was also revealed to be enlarged. This pattern remained irrespective of the presence of macroscopic lesions. SIENA (146), an automated software for brain atrophy quantification on MRI was applied by Trivedi et al. (147), to show significantly higher brain volume decline in mild-to-severe TBI as opposed to normal controls. This software was also used to quantify longitudinal brain atrophy on MRI (T1-weighted) among severe TBI patients and revealed an association between higher rates of brain atrophy and injury severity, and that brain volume change was a better predictor of long-term functional outcome as compared to functional measures at 8 weeks post-injury (148). TBM revealed localized atrophy in the brain stem, thalamus, putamen, and WM structures like the cerebellar peduncles, internal and external capsules, CC, superior and inferior longitudinal fasciculus, and corona radiata as opposed to normal controls. A surface-based morphometry (SBM) study including DTI and high-resolution MRI revealed distinct cortical thinning, GM diffusivity, and loss of integrity in the pericortical WM in TBI patients (149). Developments in the use of computational tools for objective assessment of atrophy after mild TBI have also led to an FDA approved software. It was shown to reliably capture progressive volume loss, and demonstrated degradation in whole brain (parenchyma, CSF, WM), and regional (forebrain, cortical GM, cerebellum, brainstem) structures on T1-MRI as compared to normal controls (150). This tool was subsequently shown to be more sensitive to (progressive) atrophy and asymmetry detection than radiologist interpretations (151, 152). In a very interesting application of this tool to predict previous brain volumes based on current measures, the developers also showed that reliable estimates of brain volumes were obtained on normal subjects. And that TBI patients show rapid progression of cortical atrophy, as opposed to enlargement in the subcortical nuclei and infratentorial spaces, in the few months after injury. As in the case of distinguishing NPH from AD and PD, we found that computational methods to assess atrophy and hydrocephalus cTBI are predominantly MRI based, and deep learning based approaches are lacking.

There are abundant examinations into features that can classify between NPH and asymptomatic controls, but those that predict shunt response in symptomatic NPH have not been established. Imaging studies have typically focused on predicting probable NPH due to the apparently high clinical value. Therefore, prediction of tap-test/lumbar drain responses using structural neuroimaging markers as a non-invasive alternative for a preliminary assessment of shunt response is a well-studied topic. CSF space morphological features, characterizing the narrowing of the HC/ML sulci and higher SS, enlargement of the SF and lower SS, and vertical expansion of the LVs have been shown to be highly discriminative in predicting tap-test response. The EI-z and DESH capture them in a straightforward and interpretable manner, but the latter would benefit from a computational formulation. Benedetto et al. (35) proposed the SILVER index as a potential solution. However, localized measurements of SS require manual annotation which inevitably introduces observer variability and higher computational cost. Features characterizing GM and WM integrity have also been shown to correlate with tap-test response, but nonuniformity in scanner calibration, noise, and poor soft tissue resolution may impact such density-based measures on CT scans, which is additionally limited by the need for manual annotation on both CT and MRI scans for localized anatomical measurements.

While discriminating the tap-test response is important, the low sensitivity of the tap-test itself toward shunt response (153) suggests that features optimized to predict tap-test response might reflect or exaggerate the low sensitivity. The performance reports of imaging markers as predictors of shunt response in NPH have also been mixed. Features like the iNPH Radscale have shown prognostic value in discriminating shunt responsive from healthy individuals, but there is no compelling evidence that it can predict shunt response in symptomatic individuals. Additionally, response to shunting may depend upon many factors like post-surgery care, presence of comorbidities, physical therapy, shunt complications etc. which need to be considered in future studies. Optimizing the sensitivity of radiographic evaluation to predict NPH related signal whether it is possible, probable, or definite might enable higher screening and lower cases of underdiagnosis.

Addressing the high levels of misdiagnosis in NPH is also crucial to identify and treat patients with this reversible dementia. While features for the distinction of NPH from AD have been well established, we observe that there is a lack of neuroimaging studies to distinguish NPH from PD using structural imaging. Perhaps due to heavy reliance on levodopa response in patients with suspected PD or clinical symptoms. Clinicians and researchers alike must recognize that this is a suboptimal approach which may put true NPH patients at unnecessary risk due to levodopa side effects and delay treatment with shunt-surgery. We have identified features that are discriminative of PD from controls, which may be tested for their application of distinguishing NPH from it. We encourage researchers to explore computational approaches to fully explore the potential of structural imaging in distinguishing between NPH and PD. Shunt surgery is also capable of relieving post-traumatic hydrocephalus. But research is needed to test the potential of structural imaging markers in selecting patients for surgery, as it is riddled with the problem of distinguishing between atrophy and hydrocephalus following TBI. Unfortunately, in more than 20 years post Marmarou et al.’s (13) advocacy for the use of CSF dynamics in patient selection for shunting in post-traumatic hydrocephalus, an alternative solution with noninvasive markers has not been found.

It is also important to distinguish Long Standing Overt Ventriculomegaly in Adults (LOVA), which is a chronic form of hydrocephalus, from NPH. Most cases are thought to arise due to aqueductal stenosis and may have symptomatic overlap with NPH. Patients with LOVA may also see clinical improvement with shunt surgery (154, 155). It is crucial to differentiate it from NPH as it may be extremely sensitive to pressure variations, which needs to be considered while evaluating surgical treatment (endoscopic third ventriculostomy for LOVA versus ventriculoperitoneal shunt for NPH) options. Characteristic neuroimaging finds of LOVA include (third and lateral) ventriculomegaly, decreased prominence of cortical sulci, macrocephaly, and expansion of the sella turcica due to compensatory mechanisms (156). In cases where LOVA occurs with an open aqueduct, where it may be differentiated from Late Onset Idiopathic Aqueductal Stenosis (LIAS), it becomes more crucial to distinguish it from NPH due to their clinical and symptomatic overlap (157). ICP monitoring in LOVA patients was shown to correlate with patient conditions pre- and post-surgery (158), and CSF dynamics was recommended to differentiate them from NPH patients (159). A noninvasive and accurate diagnostic score consisting of clinical features like age, presence/absence of the Hakim triad, headache, nausea/vomit, and neuroradiological features (evaluated on MRI) like the head circumference, EI-x, 3 V width, DESH, sellar bone distortion with the bulging of the 3 V floor was proposed recently by Palandri et al. (157) which was shown to classify probable NPH patients from LOVA and LIAS patients with a high AUC of 0.97, sensitivity of 95.1%, and specificity of 90.6%. Higher values of this diagnostic score correlates with higher EI-x, 3 V width, head circumference, 3 V floor bulging, and lower prominence of DESH which incorporates distinctive neuroradiological findings in LOVA as opposed to NPH. More research is needed to identify imaging markers which can differentiate shunt-responsive NPH from LOVA/LIAS.

Even though the most prominent structural impact of NPH except for WM abnormality can be visualized on CT, most research studies lean toward MRI. We identify some CT based markers which have shown good predictive performance, and further encourage researchers to not only adapt MRI-based features to CT studies, but also develop CT-specific features for this application. CT-imaging biomarkers, as affordable, timely, and accessible diagnostic solutions, with fewer contraindications than MRI may offer a hopeful prospect. Even though MRI is the preferred mode of structural imaging due to its higher soft tissue resolution as opposed to CT, the development of computational tools to extract volumetric, intensity, and texture-based CT-features may reveal the potential in characterizing structural degeneration in NPH and its differential diagnosis from AD and PD. CT is the preferred mode of evaluation for acute TBI, but the distinction between atrophy and hydrocephalus in cTBI is still heavily studied only on MRI. Advances in computational tools, particularly in machine (deep) learning and image processing, may hold the key to discovering novel CT-based markers, and objectively extracting diagnostic standard markers for these conditions.

A popular choice in semi-automatic image processing methods is to integrate domain knowledge through the volumetric characterization of brain regions that are known to be affected. While this has been successful on MRI, obtaining pixel-level ground-truth for segmentation on CT is not only expensive, but difficult to create by visual inspection. Moreover, brain regions are impacted differently in different neuropathologies and developing individual segmentation-based approaches to capture them, and their interactions can be challenging and quickly add complexity. Architectures from CNN models with the inherent capacity to learn feature representations at increasing levels of abstraction, residual networks that utilize skip connections to solve the problem of vanishing gradients in deep architectures, and the U-Net and its variants that optimally integrate local and contextual features have been limited in the assessment of NPH, and mostly on MRI. The potential of such models in solving other problems such as differentiation of (shunt-responsive) NPH from AD/PD, and atrophy/hydrocephalus in cTBI on CT scans should be tested.

Our article has a few limitations. It is a narrative review. Even though the search strategy was methodical, it was not exhaustive in terms of databases as we only included PubMed. We did not conduct a dedicated search for articles pertaining to computational methods which have been reviewed in this paper, but they were isolated from the search described in the “Search Methodology” section.

Better recognition of (normal pressure) hydrocephalus is paramount for the timely management of patients with potentially reversible dementia or brain injury. Despite a plethora of knowledge on the discriminative anatomical markers of NPH, only a few like the EI-x, CA, and DESH have notably been included in diagnostic guidelines. While prediction of tap-test/shunt response may offer most assistance to clinicians, discriminating hydrocephalic pathology from irreversible atrophy is necessary to identify patients who may be surgically treated. Anatomical markers that can accurately predict shunt surgical response in symptomatic NPH are yet to be reliably established, which may be complicated by amount of pathological progression prior to surgery, comorbidities, and post-surgical care. Longitudinal evaluation of anatomical markers correlated with symptomatic progression before and after surgery may shed light on expected recovery and planning treatment options. Additionally, there is a clear under-utilization of CT based markers in distinguishing NPH from AD, and an overall lack of studies assessing its structural differentiation from PD. Investigations into the distinction of the chronic effects of TBI, namely atrophy and (normal pressure) hydrocephalus, using structural imaging markers is also a crucially unaddressed area which may help to identify patients whose symptoms may be alleviated with shunt placement. Computational tools like image processing may help with objective measurement of features that are correlated with NPH pathology on CT; and advances in deep learning may also highlight explainable features with potential for accurate diagnosis. In emergency settings, smaller community care centers, and hospitals that may not have access to advanced imaging and the expertise to assess them, automatic methods for assessing CT scans for the accurate detection of NPH or post-traumatic hydrocephalus may be of immense value in recognizing patients with reversible symptoms. Through this effort, we urge researchers to untangle the web connecting these neurodegenerative conditions, offering hope to millions, and potentially preventing countless more from stumbling down the treacherous path of structural neurodegeneration.

SKS: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Validation, Visualization, Writing – original draft, Writing – review & editing. JD: Conceptualization, Methodology, Project administration, Writing – original draft. RG: Conceptualization, Methodology, Writing – review & editing. SC: Investigation, Writing – original draft. RK: Methodology, Project administration, Resources, Supervision, Writing – review & editing. US: Conceptualization, Investigation, Methodology, Project administration, Resources, Supervision, Writing – review & editing.