Assessment of the low-field magnetic resonance imaging for the brain scan imaging of the infant hydrocephalus

Introduction

Childhood hydrocephalus is the most frequent pediatric illness needing neurosurgery globally, with an estimated 400,000 new cases each year.[1] The majority of incidents, more than 90%, occur in LMICs, or low-and middle-income countries.[1] The disease affects over 180,000 infants in sub-Saharan Africa each year.[2-4] In newly born with hydrocephalus, the head enlarges due to an accumulation of intracranial cerebrospinal fluid (CSF). To arrange surgery for these infants, intracranial imaging is necessary. Knowing where the CSF is located in the brain and how many loculated compartments where fluid is trapped is crucial for planning surgery. It is necessary to have imaging technology that can accurately depict the contrast between the brain and CSF. We earlier suggested that a voxel size approaching 100 mm³ (eg. 3 × 3 × 10 mm³) could be enough to set treatment plans.[1]

The soft tissue and fluid that are contained in the skull can only be imaged to a limited extent on the brain. The acoustical windows of the fontanels are only open during the 1^st year of life due to skull fusion; for this much, ultrasound is only effective. The risks associated with ionizing contamination from computed tomography (CT) scans for infants can vary depending on factors such as the dose of radiation, frequency of exposure, and the infant’s age. Infants who undergo CT scans may be at higher risk of radiation-related complications due to their increased sensitivity to radiation, as their organs and tissues are still developing. The potential dangers may include an increased risk of developing cancer later in life, potential damage to developing organs, and an increased risk of other long-term health effects. Thus, the ionizing contamination associated with CT poses very high dangers to infants;[5] though, in the subarea of Saharan Africa, CT is more dominant than magnetic resonance imaging (MRI) (according to a 2011 report by World Health Organization) due to its cheap cost. Although MRI is the gold standard for pediatric neuroimaging, high-field cryogenic systems are unaffordable for the majority of the developing world due to their high cost, stringent siting requirements, and demanding maintenance schedule;[6,7] (World Health Organization, 2011).

Uganda has 0.45 CT machines per million people and only 0.08 MRI machines, based on a World Health Organization baseline country study on medical equipment done in 2014. In comparison with a financially strong country like the Netherlands has only 12 CT as well as 12 MRI machines for a million people, almost 27 times more CT machines per million people and 150 times more MRI machines per million people. There are more new instances, with at least 10 times more per year, of hydrocephalus each year in Africa than in Europe,[1] so there is a strong clinical need for internationally sustainable diagnostic imaging technologies. According to the current research,[8] low-field MRI equipment is affordable for the poor world and has the potential to diagnose and cure conditions and manage illnesses such as hydrocephalus.

The quality of an MRI image is ultimately determined by the signal-to-noise ratio (SNR) per voxel, which reflects the balance between the signal strength and the presence of noise. Voxel sizes as small as hundreds of micrometres can be pushed by systems with higher field strengths (>1.5 Tesla), which also cause an increase in signal-to-noise.[9,10] The possible voxel size is constrained by poor signal-to-noise in low-field systems (<0.1 Tesla), which also includes higher baseline noise than is typical for physicians. There are some facts and figures which are helping the authors view the presence of comorbidities or patient factors: Some medical conditions or patient factors, such as patient motion, metallic implants, or obesity, can introduce artifacts and affect the quality of MRI images, leading to increased baseline noise. If the physicians are interpreting MRI images of patients with such comorbidities or factors, it could result in higher baseline noise in their readings. It is important to note that higher baseline noise in MRI readings can potentially impact the accuracy and reliability of the interpretations.

In above Figure 1, the primary distinctions in brain picture quality between a high-field [Figure 1a] and low-field [Figure 1b] MRI system is depicted.

Close

Figure 1

(a and b) Here is a comparison of the two high-quality images taken at the medical center of the same volunteer’s brain, one in high-field and the other in low-field

Export to PPT

The long-standing and lately expanding body of evidence showing increased image quality does not always translate into better diagnostic accuracy or patient outcome is a major factor in the adoption of low-field MRI into clinical practice.[11] For many illnesses in clinical practice, there is an image quality threshold over which no further outcome-based value may be seen.[2,12] Outcome-based value refers to a concept or approach where the value of a product, service, or intervention is assessed based on its outcomes or results, rather than solely on its inputs or outputs just like health-care products such as MRI machines. It focuses on the actual impact or effectiveness of an intervention in achieving its intended outcomes, rather than just evaluating the inputs or outputs of the intervention. Multiple sclerosis, hepatic lesions, and central nervous system disorders have all been shown to respond to 0.5 Tesla MRI as accurately as 1.5 Tesla MRI in terms of diagnosis.[13,14] In addition, it has been demonstrated that a 0.064 Tesla MRI can achieve diagnostic precision comparable to that of a 1.5 Tesla MRI for neoplasms and white matter disorders. We expected that the amount of resolution, tissue contrast, and SNR supplied by CT or high-field MRI significantly exceeds this threshold, even though the minimum picture quality needed to plan an effective hydrocephalus therapy has not yet been investigated. Super-resolution augmentation of subpar MRI images has already been carried out using a variety of machine learning-based techniques.

Although interpolation-based techniques are straightforward to use, they frequently blur because they lack prior knowledge.[15] Model-based approaches[16,17] investigate the mechanism of the stochastic in the process of generating the MRI and model it with prior knowledge, although it can be difficult to build a suitable regularization for the model. The benefit of learning-based approaches is that they can model and learn how to convert poor-quality images into excellent-quality images just from data.[18] Nowadays, deep learning displays outstanding work in MRI of high resolutions.[19]

In this study, we evaluate and focus on the diagnostic value of deep learning-enhanced and low-quality pictures for hydrocephalus treatment planning. We concentrate on the post-infection of newborn hydrocephalus, which is the most prevalent kind in the partial region of Saharan Africa.[19-21]

As far as it is concerned to the outside region of the LMIC, the problem is quite uncommon LMIC,[1] and CT scans are the only source of many high-resolution comparison pictures. Three senior neurosurgeons with substantial experience treating and managing hydrocephalus in low-resource settings completed an image usefulness assessment that we devised.[15,21,22]

To categorize the images, it discusses the categorization of images related to hydrocephalus treatment planning based on their resolution, noise, and contrast between brain and CSF. It mentions the use of qualitative and quantitative metrics to assess the utility or usefulness of these images. In addition, it mentions the potential for misleading changes in low-resolution images during the improvement process using machine learning.

In the context of hydrocephalus treatment planning, medical images such as MRI or CT scans are commonly used to visualize the brain and CSF. These images need to meet certain quality thresholds in terms of resolution (i.e., clarity), noise (i.e., interference or distortion), and contrast (i.e., the difference in brightness or intensity) between brain tissue and CSF to be useful for accurate diagnosis and treatment planning.

Qualitative and quantitative metrics are employed to evaluate the quality of these images. Qualitative metrics involve subjective assessments made by human experts, such as evaluating the overall image quality, presence of artifacts, and visual clarity. Quantitative metrics, on the other hand, involve objective measurements, such as SNR, contrast-to-noise ratio (CNR), or structural similarity index, which provide quantitative measures of image quality.

The statement also mentions the potential for misleading changes in low-resolution images during the improvement process using machine learning. This refers to the use of machine learning techniques, such as image enhancement algorithms, to improve the quality of low-resolution images. While these algorithms can enhance the visual appearance of low-resolution images, they may also introduce artifacts or distortions that can be misleading, affecting the accuracy of diagnosis or treatment planning. Therefore, it is important to carefully validate the results of machine learning-based image enhancement methods to ensure that they do not lead to misleading changes in the images.

Methods

As participants in the image utility assessment, three skilled pediatric neurosurgeons with experience in LMIC and expertise in the interpretation of images of African newborns with post-infectious hydrocephalus were chosen. The CURE Children’s Hospital of Uganda treated 90 individuals with post-infectious hydrocephalus, with a median age of 3.1 months and a female gender percentage of 39%.

Each patient’s center-most image slice was chosen on a random basis from the test image (10 randomly selected as described in [Figure 2]) or an image collection from the learning library for the assessment (remaining 80). The collection of images has a resolution of 0.4 mm and is 512 × 512 (20.48 cm field of view). It is 5 mm thick on each slice.

Close

Figure 2

(a-e) The image parameter space describing all possible combinations of noise, the contrast between brain and CSF, and image resolution are visualized. There is likely to be a region of parameter combinations yielding images which are useful for hydrocephalus treatment planning (green volume), a region of parameter combinations that are not useful (pink volume), and a region of uncertainty in between (orange volume)

Export to PPT

The resolution, noise, and contrast between the brain and CSF in the ten test images all declined. The resolution was changed by changing the image’s matrix size because the field of view was consistent throughout all photos. Due to this connection, in the current work, we use the term “resolution” to refer to variations in image matrix size. As seen in Figure 2, an image parameter space was created using the following variables: There are 1,600 different parameter combinations based on the following factors: contrast reduction in the image (20 levels between 0 and 1) then having the resolution (32 × 32, 64 × 64, 128 × 128, 512 × 512), and finally the addition of noise (20 levels between 0 and 1).

Bilinear interpolation was used to downscale the resolution from the original 512 × 512 images. Bi-linear interpolation’s pixel averaging can be viewed as an approximation of the partial volume effect.

A method created especially for this task, histogram compression, was used to lessen the contrast between the brain and CSF. To simulate a reduction in tissue contrast, histogram compression iteratively reduces the brain and CSF’s histogram of gray-scale values into a narrower gray-scale bandwidth. In accordance with the known noise properties of CT scans, Gaussian noise with a mean equal to variance was applied.[23]

To depict both useful and useless images, the noise produced was scaled by clinical inspection at each resolution because lower-resolution images are more sensitive to noise. The noise variance added was scaled by resolution as follows and normalized to the maximum value: from 0 to 0.001 (32 × 32), 0 to 0.01 (64 × 64), 0 to 0.05 (128 × 128), and 0 to 0.13 (512 × 512).

These papers having the idea used deep learning networks to improve low-quality images by utilizing low-rank structural prior knowledge.[24-26] Using this work as a foundation, we created a deep learning network that can concurrently enhance and segment data CT pictures of infants with intentionally deteriorated hydrocephalus.

To improve low-quality CT images, a single encoder dual decoder architecture was applied, which was modeled after the DenseNet network discussed in.[11,27] Deep learning networks were trained using library pictures at two resolutions (64 64 and 128 128) at seven locations in parameter space, as shown in Figure 2. Networks were trained for both resolutions at (1) (0.3, 0.3), (2) (0.6, 0.3), (3) (0.3, 0.6), (4) (0.6, 0.6), (5) (0.9, 0.6), (6) (0.6, 0.9), and (7) (0.9, 0.9) with noise introduced as the x-coordinate and contrast reduction as the y-coordinate. Network 1 is the least damaged. At each of the 14 network locations, 80 library images were degraded to build the networks, which were then trained using the original, undeleted image as the ground truth. Following training, the network locations degraded the 10 test images and enhanced them, creating 140 deep learning-enhanced images.

The panel of experts received all 140 deep learning-enhanced photographs and 420 cases at random from the 1,600 parameter combinations applied to the 10 test images. The evaluation of the image’s utility was split into two sections. 140 panels of four photos each were used to display the images in Part 1 as illustrated in Figure 3a. One improved image location and three degraded image locations were randomly chosen for each of the 140 panels. The existence of improved photographs was not disclosed to the expert. The expert in each panel was asked to decide which, if any, of the four photos was clinically helpful for formulating a treatment strategy for hydrocephalus (see Supplementary Methods for full instructions). By summing the scores for each image to be classed as useful, uncertain, or not useful, the data from the three experts were integrated. An image scored a 3 if it was deemed beneficial by all three experts (i.e., Useful). 0 was given to an image if all experts concurred that it was not used (i.e., Not Useful). Images that were unclear either scored a 1 or a 2.

Close

Figure 3

(a-i) The figure requires the results of the assessment. In the figure, we show an example panel from part 1 of the assessment. The lower left image is an enhanced image and all other images are degraded. The experts must indicate which (if any) is useful

Export to PPT

As seen in Figure 4a from Part 2, the experts’ updated images were displayed next to their corresponding 512 × 512 non-degraded counterparts. The improved version’s spatial flaws were assessed by the experts to see if they were acceptable or if they would alter treatment suggestions (see Supplementary Methods for full instructions). Moreover, Part 2 data were combined using score addition. Part 2 enhanced photographs were graded as useful (meaning they were useful in both Parts 1 and 2) if they received a 3, doubtful if they received a 1 or 2, and misleading if they received a 0. (i.e., useful in Part 1, but shown to have an unacceptable error in Part 2).

Close

Figure 4

(a-d) The figure shows the results of the assessment. The left column of images is the ground truth and the right column is the enhanced versions. It also shows the predicted usefulness likelihood of the enhanced images based on CNR after enhancement. We compare the usefulness and likelihood of the degraded images

Export to PPT

In addition, an analysis of inter-rater reliability was performed using a variation of Cohen’s Kappa, as defined by another scholar,[28] which compensates for the presence of prevalence in the data as well as evaluator bias (Supplementary Methods). The data were separated into three sections for this analysis: (1) classification of Part 1 degraded photos, (2) classification of Part 1 enhanced images, and (3) classification of Part 2 enhanced images. The Byrt-Kappa statistic[29-31] was calculated for all possible evaluator pairings, and inferences about the agreement were formed based on the interpretation of Kappa values as recommended by Byrt and Hallgren.[32,33]

To evaluate the capacity of contrast, noise, and CNR to predict image categorization, univariate and multivariate logistic regression were used. The logistic regression models’ quality of fit was evaluated using a deviance statistic. The model deviance is a Chi-squared statistic that evaluates the difference between the maximum log-likelihood of the chosen model and the null model (i.e., the average probability of classification at a given resolution being useful).

Results

Part 1. Image displaying a list of features that make an image useful.

First, we shall characterize the link between contrast, noise, resolution, and very lastly usefulness. There is a good similarity between examiner 1 and 2 (K = 0.33), a strong agreement between examiner 2 and 3 (K = 0.94), and a fair agreement between examiner 1 and 3 (K = 0.36), as determined by the interrater reliability for the classification of degraded images. There was a significant prevalence of Part 1 improved images being categorized as beneficial by all three judges (see Supplementary results). As such, inter-rater reliability calculations for these data are not informative and all evaluators are in near-perfect agreement. In Figure 3a, we display several deteriorated images, the lower left of which has undergone deep learning enhancement. The contrast and noise of each image are related to the 64 × 64 resolution image categorization decisions in the left panel of Figure 3b. The continuous CNR between the brain and CSF is depicted in Figure 3b by the solid contour lines, which are averaged throughout the whole collection of pictures. Contrast and noise were used as predictors in the multivariate logistic regression model that produced the dotted lines, which demonstrate constant usefulness likelihood. The multivariate logistic regression model offered a significant fit with p-values <0.01 for images with each of the four resolutions taken into consideration (p32 × 32 = 7e-6, p64 × 64 = 4e-27, p128 × 128 = 2e-17, and p512 × 512 = 8e-32).

The lines and the average contrast-noise contours indicate a constant similarity that the image is qualitatively agreed upon as well as useful, as can be seen. With areas under their curves >0.85 (curves for the whole dataset in [Figure 3]), average contrast-to-noise and likelihood are both relatively effective classifiers of image utility, as indicated by Figure 3 receiver operating characteristic curves in the right figure.

Figure 3c demonstrates that individual image contrast-to-noise alone is a substantial predictor of usefulness likelihood, stratified by resolution, as average contrast-to-noise appears to be a successful classifier. Based on its CNR = 13, the gray circle in Figure 1a depicts the possibility that the 256 × 256 brain image from the 3 Tesla system will be usable. In Figure 1b gray diamond (CNR = 4) for the 128 × 128 brain image from the 0.05 Tesla system, the identical result is displayed. Despite having twice the resolution and 3 times the CNR as the image created by the 3T system, both images have a predicted utility likelihood of one. The four inset panels of Figure 3c display the raw categorization data from Part 1 for each resolution. The logistic regression model is represented by solid lines, whereas the 95% confidence intervals around the fit are represented by dashed lines.

Part 2: Reconstruction of errors is acceptable

We then look into how deep learning enhancement affects image classification. There is only a little agreement between evaluators 1 and 2 (K = 0.15), and only moderate agreement between evaluators 1 and 3 (K = 0.48), according to the inter-rater reliability for the classification of improved images in Part 2. It is a side-by-side comparison of the facts (left column) and related improved photographs, as seen in Figure 4a (right column). Take note of the minor inaccuracies in the top-right image’s placement of the brain and the CSF as well as the more obvious flaws in the lower-right image. Despite these spatial flaws, the enhancement network greatly raises CNR, as seen in the plot in Figure. The average CNR of test images at each network point is shown before and after augmentation in Figure 4b using the logistic models developed in Part 1 of this article. Based on higher CNR, these Preda ICT data have a very high possibility of being useful for enhanced photos. Although the classification of enhanced images in Part 1 did closely resemble the expectation of high usefulness, the likelihood, the table in Figure 4c demonstrates that many improved images have defects that are not clinically acceptable. For these images, we additionally classify them as misleading (i.e. images that were deemed Useful in Part 1 but had unacceptable errors in Part 2).

A new logistic regression model with pre-enhancement noise and image contrast as predictors was created for Part 2 because the logistic models generated in Part 1 do not adequately represent the Part 2 categorization. Only contrast showed significance [Figures 5 and 6] so, to observe the practicality probability of a noisy image (as in Part 1) with the danger of misleading mistakes in a boosted image (as in Part 2), the noise was eliminated from the motor. Based on Part 2 classification, an extra logistic regression model was calculated with CNR as the predictor [Figure 4d]. Based on a univariate logistic regression with the usefulness likelihood of the enhanced images as the predictor, the risk of misleading results is computed to be 1 minus the likelihood that the upgraded images will be beneficial. A 128 × 128 image is more likely to be helpful when it is degraded (left vertical axis) and probably to be less misleading when it is improved as CNR rises (right vertical axis). It should be noted that no CNR value exists, for which there is a low possibility of the poor-quality image being usable and a low danger of producing an inaccurate image by augmentation.

Close

Figure 5

We can see the role of the histogram and its direct impression on our brain scan images

Export to PPT

Close

Figure 6

The figure shows the various resolution images and their impact on CNR and LogReg images. So, we could say that the image’s resolution is directly proportional to image sensitivity

Export to PPT

Discussion

Conclusion

Instead of its esthetic appeal, a clinical medical image’s genuine worth lies in the information; it provides on how to treat patients and in the outcomes that follow. We have demonstrated that treatment planning for hydrocephalus can benefit from using lower-quality pictures that are not often regarded as acceptable. In addition, the chance of a meaningful image for hydrocephalus treatment planning is predicted by the image resolution and CNR of the brain and CSF. Although deep learning can significantly improve an image’s visual quality despite its severe degradation, there is a significant risk of inaccurate outcomes, therefore algorithmic standards should be devised to prevent structural changes that could be dangerous for clinical interpretation. Currently, the least enhanced low-resolution images that keep the structural details intact and are not distorted by excessive deep learning processing may be the most valuable; in fact, new low-field MRI technologies are capable of producing useful images for hydrocephalus treatment planning without enhancement. Our findings call for a re-evaluation of what constitutes acceptable image quality for clinical usage as well as new guidelines for evaluating the cost-effectiveness of sustainable imaging technologies that can increase access to diagnostic imaging globally.

However, like other new technologies, this technology also has some pros, as well as cons and those points, should be considered according to the seriousness of the patient, need for this technology, and also working background also. Pros of Assessing the Low-Field magnetic resonance imaging (MRI) for the Brain Scan Image of the Infant Hydrocephalus are Low-field MRI is a non-invasive imaging modality that does not require ionizing radiation, making it a safer option for infants with hydrocephalus compared to other imaging techniques such as CT scans; MRI provides high soft-tissue contrast, allowing for the detailed visualization of brain structures, which is particularly beneficial for assessing the brain of infants with hydrocephalus where precise anatomical information is crucial for diagnosis and treatment planning; low-field MRI has the potential for the early diagnosis of hydrocephalus in infants, allowing for prompt medical intervention and management, which can prevent further complications and improve outcomes. On the other hand, some drawbacks are also connected to this technology. Some of them as follows: Low-field MRI typically has lower image quality compared to high-field MRI, which may result in reduced spatial resolution, SNR, and CNR. This may affect the ability to accurately identify and characterize hydrocephalus in infants, especially in cases with subtle findings; low-field MRI may require longer scan times compared to high-field MRI, which can be challenging in infants due to their limited ability to cooperate and remain still during the imaging procedure; In cases of complex hydrocephalus, such as when there are associated brain abnormalities or when evaluating shunt function, low-field MRI may have limitations in providing detailed and accurate information compared to high-field MRI, which may have higher diagnostic accuracy; It’s important to note that the pros and cons of assessing low-field MRI for infant hydrocephalus may vary depending on the specific imaging protocol, equipment, and expertise of the imaging team. It’s essential to carefully consider the benefits and limitations of low-field MRI in the context of each individual case and consult with a qualified medical professional for proper evaluation and management of hydrocephalus in infants.

Ethics Approval and Consent to Participate

This study was conducted in accordance with the ethical principles of the declaration. The Institutional Ethics Committee approval was obtained before the study and registered at Clinical Trial Registry India. Written informed consent was obtained from all the participants.

Availability of Data and Material

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Competing Interests

The authors declare no conflicts of interest.

Funding Statement

No external funding was received for this project. It is an investigator-initiated academic study.

Authors’ Contribution

SS: Performed data collection and manuscript drafting. QMR: Designed the study, conceptualization, methodology, and manuscript drafting. Both authors reviewed and approved the final version of the manuscript.