Assessing subtle brain injury related to cognitive impairment in obese pediatric obstructive sleep apnea using 3D MRI texture analysis Authors: Sina Tafreshi, Daniel Sare, Alex Rus, Indra Narang, Andrea Kassner Published: 2026 ABSTRACT AbstractObstructive sleep apnea (OSA) is a frequent complication of childhood obesity associated with neurocognitive impairment. Subtle brain injury caused by intermittent hypoxia and sleep fragmentation has been implicated, yet cannot be detected by structural MRI. This study investigated whether three-dimensional texture analysis (3D TA) of T2-weighted MRI could detect subtle microstructural alterations in brain regions affected by cognitive impairment in obese children with moderate to severe OSA compared to healthy age-matched controls. Texture features were extracted, ranked using the Fisher Criterion, and classified with k-nearest neighbors across regions of interest. Among 23 children with OSA and 26 healthy controls, classification accuracy was highest in the left amygdala (87%), right hippocampus (82%), and left insula (82%), with intensity-based features proving most discriminative. In two patients treated with continuous positive airway pressure (CPAP), follow-up scans indicated normalization of brain texture features. These results highlight 3D TA as a sensitive, clinically accessible method for detecting subtle brain changes in pediatric obesity related OSA. Incorporating this approach into clinical practice may improve treatment monitoring and outcomes, though validation in larger cohorts is needed to confirm its validity and to further assess how these markers change with therapies such as CPAP.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-56328-5.Subject terms: Diseases, Medical research, Neurology, Neuroscience Introduction Obesity is the primary risk factor for obstructive sleep apnea (OSA) in children1, as excess fat deposits can lead to the collapse of the pharyngeal airway and restriction of the thoracic wall2. However, in younger children, OSA is often associated with enlarged adenoids and tonsils2. OSA is the most common sleep-related breathing disorder which affects around 30% of obese children3,4 and 2% of the general pediatric population5. This disorder is characterized by intermittent narrowing and collapse of the airway during sleep, resulting in recurrent hypoxic events, elevated CO2 levels, and sleep fragmentation6,7. These manifestations can lead to a range of symptoms including daytime sleepiness and headaches. More importantly, OSA in children is associated with increased risk of cardiovascular, metabolic, cognitive, and behavioral comorbidities8–12, underscoring the importance of early detection and intervention given their developmental stage. The effects of OSA on cognitive and behavioral functioning in children are especially concerning, as they can adversely impact healthy development and contribute to poor academic performance13,14. Obese and non-obese children with OSA often exhibit cognitive impairments in attention, memory, executive function, impulsivity, and emotional regulation15,16. Studies have also shown that children with OSA frequently experience symptoms of depression and anxiety17,18. Notably, certain neurocognitive impairments can show improvement with surgery and/or appropriate continuous positive airway pressure (CPAP) therapy19, highlighting the reversibility of these abnormalities and supporting a causal relationship between OSA and neurocognitive dysfunction. These cognitive and behavioral abnormalities are presumably caused by subtle brain injury as a result of the physiologic disturbances associated with OSA. Among these disturbances, recurrent hypoxic events play a key role by triggering systemic inflammation, leading to neuronal damage and endothelial dysfunction that compromise vascular integrity and brain health20–22. Elevated levels of oxidative stress associated with OSA, to which the developing brain is particularly susceptible, exacerbates this damage23,24. Additionally, sleep fragmentation has been shown to independently affect neurocognitive functioning in children with OSA12, indicating its role in contributing to subtle brain injuries. These mechanisms may also impair neurogenesis25, hindering the brain’s ability to generate new neurons and repair damaged tissue, thereby impacting brain development. Several brain regions can be impacted due to the presence of OSA in association with observed cognitive and behavioral impairments. The emotional dysregulation and social difficulties observed in obese children with OSA can be linked to the amygdala and insula, which are regions vital for emotion processing and social cognition26,27. The impairments in memory and learning can be linked to the hippocampus, which is a region responsible for memory formation and retrieval28. Lastly, the impairments in attention, executive function, and impulsivity can be linked to the orbitofrontal cortex (OFC) and anterior cingulate cortex (ACC), which are regions involved in higher-order cognitive processing and behavior regulation29,30. These five brain regions are central to the cognitive and behavioral functions disrupted by OSA, emphasizing their potential importance in understanding the widespread impact of OSA on the developing brain. These anatomical correlates of cognitive impairment have been extensively explored in adults with OSA through neuroimaging studies, with magnetic resonance imaging (MRI) being a valuable tool for detecting changes in brain structure, function, and tissue integrity. In obese and non-obese children with OSA, structural brain MRI findings show considerable variability, although alterations in gray matter volume and cortical thickness have been observed in the aforementioned brain regions31–35. Regarding functional disruptions, Kheirandish-Gozal et al. used functional MRI (fMRI) in non-obese children with OSA and found reduced activity in the ACC during an executive function task and in the amygdala during an empathy task36, highlighting functional abnormalities in regions critical for executive function and emotional processing. In a meta-analysis of structural and functional brain alterations in obese and non-obese adults with OSA, Huang et al. found the most significant decrease in gray matter volume in the OFC, along with decreases in the ACC and amygdala37. Additionally, they reported altered brain activity in the ACC, amygdala, and insula. Disruptions in brain tissue integrity have also been reported, as changes in mean diffusivity values have been observed in the hippocampus, OFC, and insula in obese and non-obese children with OSA38,39. Furthermore, Halbower et al. used MR spectroscopy to assess brain tissue integrity in mostly non-obese children with OSA and reported reduced metabolite ratios in the hippocampus and OFC40. These documented structural and functional brain abnormalities collectively underscore the pathophysiological sequelae of OSA that leads to subtle brain injury. While these subtle brain injuries have been documented in the research setting, their identification in clinical practice remains a significant challenge. Many of the neuroimaging techniques used to detect these subtle brain injuries (e.g. fMRI, diffusion tensor imaging, MR spectroscopy) are not commonly performed in the clinical setting due to their complexity, cost, and limited accessibility. Although standard structural MRI is more widely available and performed in the clinical setting, it is still not commonly used in the clinical assessment of children with OSA, as the brain abnormalities associated with OSA are often too subtle to be identified through conventional radiological assessment. This highlights the need for more sensitive methods that can be applied to standard clinical imaging to detect subtle brain injury associated with OSA that might otherwise go unnoticed. Given these challenges, texture analysis offers a promising approach as a tool that can uncover microstructural changes on standard structural brain MRI by quantifying signal intensity patterns using first- and second-order statistical texture features, offering a more sensitive approach for identifying subtle brain abnormalities. In particular, the second-order texture features have the advantage of providing information on the spatial relationship of voxel pairs, making it sensitive to subtle signal intensity variations on brain MRI that cannot be visually detected41. The process of texture analysis involves three key steps: feature extraction, feature selection, and feature classification. Feature extraction involves calculating the texture features from either 2D or 3D MRI data. However, compared to 2D, 3D texture analysis offers greater spatial detail by examining voxel relationships across all dimensions. Feature selection then identifies the most informative texture features for distinguishing between groups to reduce dimensionality and improve classification accuracy. Advanced feature selection techniques, such as the Fisher Criterion, offer a robust approach that make them particularly effective for high dimensionality datasets42. Finally, feature classification applies machine learning or statistical methods to classify patterns associated with different clinical outcomes, analogous to a radiologist’s clinical interpretation. Among the various classification methods, the k-nearest-neighbors (k-NN) algorithm, a supervised machine learning technique, often achieves the highest classification accuracy when compared to similar machine learning techniques43. This process of texture analysis is applicable to T1- and T2-weighted MRI data; however, T2-weighted MRI data is particularly well-suited for detecting subtle brain injury due to its high sensitivity to variations in tissue composition and water content. 3D texture analysis has been previously applied in other studies to detect subtle brain abnormalities in diseases such as epilepsy44–46, multiple sclerosis47–49, and traumatic brain injury50,51. Similarly, Kheirandish-Gozal et al. successfully used 3D MRI texture analysis to detect subtle brain injury in children with OSA52. However, their study included only non-obese children and was limited to a single first-order texture feature (entropy), thereby overlooking the use of advanced feature selection and classification methods to investigate the wider range of other texture features that could potentially provide more detailed information on subtle brain injury. The aim of this study was to apply 3D texture analysis, incorporating advanced feature selection using the Fisher Criterion and supervised machine learning classification using the k-NN algorithm, to quantify and compare texture features obtained from T2-weighted structural MRI data in regions of interest (ROIs) affected by cognitive impairment (OFC, hippocampi, amygdalae, ACC, and insulae as reported in the literature) in obese children with and without OSA. We hypothesized that this approach can reveal subtle structural brain abnormalities that may not be visible with standard visual inspection in obese children with OSA compared to obese children with no OSA and that these abnormalities would potentially reverse with CPAP therapy at follow-up MRI. Methods This is a prospective study. Subject recruitment, anthropometric measurements, clinical assessments, and MRI acquisition were all performed at the Hospital for Sick Children in Toronto in accordance with all relevant guidelines and regulations. Informed consent was obtained from all individual participants included in the study and/or their legal guardian(s). All study protocols were approved by the institutional Research Ethics Board (REB Number: 1000050597). Subjects Obese participants aged between 8 and 18 years were prospectively recruited for this study. Inclusion criteria from the study were BMI over the 95th percentile for age and sex, based on the CDC growth chart for the year 2000, and informed consent by the subjects and their guardians. Exclusion criteria included underlying neurological problems (including a history of stroke or developmental delay), co-existing central sleep apnea, hypertension, obesity secondary to other diseases and syndromes (e.g., Prader-Willi syndrome), pregnancy, recent infections or sickness, adenotonsillar hypertrophy amenable to surgical intervention, severe respiratory distress, or history of smoking/illicit drug use. Polysomnography The subjects underwent a standard overnight polysomnography (Natus Sleepworks, Natus Medical Incorporates, San Carlos, CA, USA), based on the American Academy of Sleep Medicine guidelines53. All PSG exams were scored according to the American Academy of Sleep medicine scoring guidelines54 by a registered certified polysomnographic technician. All sleep studies were reviewed and interpreted by an experienced pediatric sleep physician (IN). The obstructive apnea-hypopnea index (OAHI) was used to diagnose OSA based on the following criteria: the number of obstructive apneas, mixed apneas, and obstructive hypopneas per hour of sleep. Based on the number of events from the OAHI, the subjects were divided into two groups: No OSA (OAHI < 1.5 events per hour) and Moderate/Severe OSA (OAHI ≥ 5 events per hour)54–57. Two of the Moderate/Severe OSA subjects underwent CPAP therapy and received follow-up MRI scans. MRI data acquisition Imaging data were acquired on a 3 T MRI system (Siemens Magnetom Trio, Siemens Medical Solutions, Erlangen, Germany) using a 32-channel head coil. Anatomical images were acquired using a T1-weighted MPRAGE sequence (TR/TE = 2300/2.98 ms, TI = 900ms, FOV=256 mm, voxel size = 1 × 1 × 1 mm, FA = 9°). Axial T2-weighted images were acquired using a fluid-attenuated inversion-recovery (FLAIR) sequence (TR/TE = 9000/85 ms, FOV = 220 mm, slices = 25, slice thickness = 4.5 mm). MRI texture analysis All subjects underwent processing using an in-house pipeline including pre-processing, ROI selection, and feature extraction, selection, classification, and evaluation. All these steps of the pipeline are summarized on the workflow diagram in Fig. 1. Workflow of MRI texture analysis processing pipeline. The pipeline consists of preprocessing T1-weighted and T2-weighted 3 T MRI data followed by feature extraction, feature selection, classification, and evaluation for each ROI. MRI pre-processing T1-weighted anatomical images were pre-processed and segmented using the Recon-All pipeline in Freesurfer – Version 7.3.2 (Athinoula A. Martinos Center for Biomedical Imaging, Harvard-MIT, Boston) using the Aseg and Desikan-Killany-Tourville atlases. Subcortical regions related to cognitive impairment for each hemisphere included the orbitofrontal cortex, hippocampus, amygdala, anterior cingulate cortex, and insula. These were manually reviewed and corrected when needed, then exported as ROIs. T2-weighted FLAIR images were brain extracted and co-registered to the T1-anatomical images using FSL (FMRIB, Oxford, UK). Feature extraction ROIs and the calculation of texture features were performed by using LIFEx – Version 7.3.4 (IMIV, France). ROIs were mapped onto the co-registered T2-weighted images in LIFEx and 3D texture analysis was performed extracting 139 combined first and second-order texture features per ROI. Feature selection, classification, and model evaluation in this study employ supervised learning. The methodology identifies the most discriminatory first and second-order texture features to highlight significant cognitive ROIs and features in the brains of obese children with pediatric obstructive sleep apnea. The analysis was performed primarily using Scikit-Learn, an open-source, commercially available, machine learning and data mining framework for Python developed in 201058. An extended description for feature selection, classification, and model evaluation methods can be found in Supplemental Materials. Feature selection Feature selection was performed using the Fisher Criterion. This technique ranks features based on their ability to maximize inter-class variance while minimizing intra-class variance using the Fisher Score. The Fisher Score was computed for each feature, and the top 10 ranked features were selected to create a reduced-dimensionality dataset specific to each cognitive ROI. Prior to Fisher Score computation, all features were standardized to have zero mean and unit variance, ensuring that features with larger numerical ranges did not disproportionately influence the feature selection process. Classification and evaluation Following feature selection, classification was performed using the k-nearest neighbors (K-NN) algorithm. K-NN is a non-parametric, instance-based machine learning method commonly used for binary classification tasks; thus, it was selected for its ability to handle high-dimensional data and its adaptability to small sample sizes. To ensure model stability and minimize the risk of overfitting, a 5-fold cross-validation strategy was utilized to assess classification performance. Within this framework, hyperparameter tuning was conducted by iteratively evaluating different values of k and selecting the value that produced the highest classification accuracy. Classification accuracy metrics were used to assess the significance of how well the selected features could differentiate subtle brain abnormalities between obese subjects with and without OSA. The subcortical ROIs related to cognitive impairment and texture features in anatomical areas most relevant to cognitive impairments in pediatric obese OSA are identified based on their ability to differentiate between groups. By leveraging classification accuracy metrics statistical significance of these features can be assessed. CPAP classification The same algorithm was used to attempt to predict the condition of two patients who, when originally classified as Moderate/Severe OSA, underwent CPAP treatment, and subsequently a follow up scan. The rest of the subject data (Moderate/Severe and healthy controls) was used to train the model, but instead of evaluating the accuracy of the model on bootstrap iterations, the two patients were used as the ‘test set”, where they were classified by ROI as either Moderate/Severe OSA or No OSA. Statistical analysis To evaluate the clinical relevance of the texture features, Spearman’s rank correlation was used to assess the relationship between the top-ranked texture features per ROI and PSG markers. Statistical significance was defined as p ¸0.05, with all analyses performed using the SciPy library in Python. Results Subject demographics The demographic and clinical data of the subjects is shown in Table 1. In brief, 23 subjects with Moderate-Severe OSA and 26 healthy controls were recruited and completed imaging. There were no significant differences between groups in age (p = 0.334) and BMI (p = 0.779), while there were significantly higher OAHI scores in the Moderate/Severe OSA group compared to the control group (p < 0.001). There was also a significantly higher proportion of males in the Moderate/Severe OSA group compared to the control group (p = 0.021). Table 1Demographic data and Polysomnography measures of patients with Moderate/Severe OSA and No OSA (mean ± standard deviation).VariablesNo OSA (n = 26)Moderate/Severe OSA (n = 23)P-valueAge (Years)14.53 ± 2.6913.83 ± 2.420.334Sex (Male: Female)10:1617:60.021BMI z-score3.30 ± 0.993.22 ± 0.860.779OAHI1.58 ± 1.5221.41 ± 23.88< 0.001Arousal Index (per hour)9.72 ± 5.5717.02 ± 9.770.011Desaturation Index (per hour)2.77 ± 3.5718.22 ± 21.220.020Total Sleep Time (min)346.03 ± 73.01369.05 ± 81.820.359Sleep Efficiency (%)79.56 ± 11.4580.33 ± 15.280.854Mean Sleep SaO2 (%)96.79 ± 1.0796.61 ± 1.070.592Wake Nadir SaO2 (%)91.58 ± 5.4288.77 ± 4.810.097Sleep Nadir SaO2 (%)91.08 ± 1.8987.07 ± 4.45< 0.001Time Below 90% Saturation (%)0.025 ± 0.0720.47 ± 0.650.004 Demographic data and Polysomnography measures of patients with Moderate/Severe OSA and No OSA (mean ± standard deviation). Classification performance The classification performance of texture analysis using the l-NN algorithm with bootstrap validation demonstrated varying levels of accuracy across all 10 ROIs in determining significant discriminatory texture feature difference in Moderate/Severe OSA compared to controls (Table 2). Among the ROIs analyzed, the left amygdala exhibited the highest classification accuracy of 87.15% (standard error (SE) = 0.00106) with a 95% confidence interval (CI) ranging from 80.24 to 93.45%. This was followed by the right hippocampus with 82.30% (SE = 0.001036, CI: 77.45–89.91%) and the left insula at 82.17% (SE = 0.001058, 65% CI: 77). Table 2Classification accuracy, standard error, and confidence interval (CI) of bootstrapped (n = 1000) k-NN model.ROIFisher – kNN (Bootstrap validation)Standard Error95% CI Lower95% CI UpperACC Left81.80%0.11%73.04%87.11%ACC Right80.03%0.13%71.53%87.84%Amygdala Left87.15%0.11%80.24%93.45%Amygdala Right73.31%0.12%71.33%86.37%Insula Left82.17%0.11%77.12%90.12%Insula Right77.43%0.11%74.33%88.08%HPC Left76.17%0.12%72.82%87.66%HPC Right82.30%0.10%77.45%89.91%OFC Left75.30%0.12%72.83%87.18%OFC Right72.58%0.13%70.04%85.82% Classification accuracy, standard error, and confidence interval (CI) of bootstrapped (n = 1000) k-NN model. Figure 2a summarizes the mean ROC AUC for each ROI in contrast to a chance curve, illustrating overall classification performance (12–90.12%). Figure 2b shows the representative ROC curves for each bootstrap iteration in the left amygdala of the Fisher analysis and k-NN classifier pipeline. Each iteration generated a unique AUC value, and the mean ROC AUC across all iterations is highlighted. All values can be seen in Table 2. (A) Mean ROC curves and computed AUC for each ROI. (B) Representative ROC curves for each bootstrap iteration in the Left Amygdala for the Fisher analysis and k-NN classifier pipeline. The legend displays sample iterations for each calculated over 1000 iterations. From the 1000 iterations, a mean ROC value is computed for each ROI. When determining significant discriminatory texture features, intensity-based features were commonly in the top 10 performing features across all ROIs, showing significance in the left and right ACC, the right amygdala, and the right insula (see Fig. 3). Violin plots showing the distribution of normalized feature values (0–1) for the top 10 performing texture features in most significant regions (A) Left amygdala, (B) left and (C) right hippocampus, and (D) left orbitofrontal cortex for healthy controls and OSA. Significant differences between healthy and controls were seen in each ROI for multiple texture features (student’s T-test, where * p < 0.05, ** p < 0.01). CPAP classification performance The two OSA patients that underwent CPAP therapy and subsequent follow-up imaging post treatment were reclassified for each ROI as either as Moderate/Severe or No OSA. A weighted majority voting algorithm was used after discovering that certain ROIs were classifying for Moderate/Severe OSA, while others were not. This approach took into consideration the amount of features in each ROI which were classifying for OSA (or No OSA) in each of the patients. The overall classification was that both CPAP patients, originally diagnosed with Moderate/Severe OSA, now classify as No OSA in these cognitive ROIs based on their respective texture features (Table 3). Table 3Post-CPAP reclassification of MRI follow-up Moderate/Severe OSA subjects using severity using weighted majority voting per ROI.ROIOSA CPAP MRI Follow-UpPatient 1 ClassificationOSA CPAP MRI Follow-UpPatient 2 ClassificationACC RightNo OSANo OSAACC LeftNo OSAModerate/Severe OSAAmygdala LeftNo OSANo OSAAmygdala RightModerate/Severe OSANo OSAInsula LeftNo OSANo OSAInsula RightNo OSAModerate/Severe OSAHPC LeftNo OSAModerate/Severe OSAHPC RightNo OSANo OSAOFC LeftNo OSAModerate/Severe OSAOFC RightNo OSAModerate/Severe OSAOverall Classification No OSA No OSA Post-CPAP reclassification of MRI follow-up Moderate/Severe OSA subjects using severity using weighted majority voting per ROI. Correlation of texture features with PSG Spearman correlation analysis revealed significant associations between discriminatory texture features and objective markers of OSA severity The Left OFC demonstrated the most robust associations with sleep apnea severity. Significant positive correlations were found between OAHI (Obstructive Apnea-Hypopnea Index) and features such as Intensity Histogram Area Under Curve (r = 0.39, p = 0.015) and Intensity Range (r = 0.38, p = 0.019). Additionally, higher Desaturation Indices were linked to increased texture complexity (e.g., GLCM Inverse Difference Moment). Conversely, the Left Hippocampus showed significant negative correlations between Wake Nadir SaO2 and intensity-based features like Total Lesion Glycolysis and Integrated Intensity (r = −0.32, p = 0.047), suggesting that diminished structural uniformity in the hippocampus is closely associated with chronic nocturnal hypoxia. While trends were observed in the Left Amygdala, they did not each statistical significance, though several trends were noted (p < 0.1). Specifically, Small Zone Low Grey Level Emphasis showed a weak positive trend with Sleep Nadir SaO2 (r = 0.29, p = 0.07), and Intensity Histogram Variance showed a negative trend with the Desaturation Index (r = −0.29, p = 0.08). Spearman correlation coefficients are displayed between the top-ranked texture features and PSG parameters across four subcortical regions: (A) Left Amygdala, (B) Left Hippocampus, (C) Right Hippocampus, and (D) Left Orbitofrontal Cortex (OFC). Notable significant clusters (p < 0.05) are observed in the Left OFC and Left Hippocampus, particularly between intensity-based heterogeneity features and markers of hypoxic burden (e.g., OAHI and SaO2 nadirs), while the Left Amygdala and Right Hippocampus demonstrate more localized or non-significant trends. Discussion This is the first study to investigate structural brain abnormalities indicative of subtle brain injury in obese children with OSA using 3D texture analysis of brain MRI. Our primary aim was to determine whether 3D texture analysis could uncover microstructural brain changes in ROIs associated with cognitive impairment in this population. Cognitive impairments in children with OSA often manifest as deficits in attention, memory, and executive function15,16. These impairments are likely a result of the pathophysiological sequelae of OSA, including intermittent hypoxia, sleep fragmentation, and inflammation12,20–24, which can lead to subtle brain injury over time. However, the lack of a clinically applicable tool and the limited sensitivity of conventional radiological assessment for detecting subtle brain injury represents a major gap that could potentially be addressed through the application of 3D texture analysis. We applied advanced texture analysis methods, including the Fisher criterion for feature selection and the k-NN algorithm for supervised classification, to analyze texture features from 10 brain ROIs associated with cognitive impairment in obese children with OSA. Our findings revealed significant discriminatory texture features between the OSA and control groups across all ROIs, suggesting that 3D texture analysis may be a sensitive and effective tool for identifying subtle brain injury in obese children with OSA. While all the brain regions that were evaluated showed significant findings, 3 regions emerged as particularly significant due to their combination of high classification accuracy, statistically significant texture features, and associations with markers of OSA severity: the amygdala, hippocampus, and OFC. These regions were prioritized for further discussions given their consistent involvement across all analytical approaches. The left amygdala showed the highest classification accuracy at 87.15% along with numerous significant texture features, while the right amygdala showed considerably lower classification accuracy at 73.31% with minimal significant texture features. Both the right and left hippocampus showed strong classification accuracies of 82.30% and 76.17%, respectively, with multiple significant texture features bilaterally. Although the OFC had lower classification accuracies (75.30% left, 72.58% right), the left OFC showed one the greatest number of significant texture features among all ROIs. The following sections examine each of these regions in detail, contextualizing our findings within existing literature on the structural brain changes associated with OSA. Amygdala The amygdala plays a key role in emotional regulation, stress response, and memory processing, which are functions that are commonly disrupted in children with OSA26. Previous neuroimaging studies have demonstrated microstructural and functional abnormalities in the amygdala among individuals with OSA. A meta-analysis of structural and functional MRI studies in adults with OSA reported both hyperactivation and reduced grey matter volume in the amygdala compared to controls37. The authors suggested that the hyperactivation of the amygdala represents a compensatory mechanism for the decreased grey matter volume. This hyperactivation is associated with increased metabolic demand which requires more blood and oxygen, resulting in relative ischemia and hypoxia that contribute to further damage over time. This process may underlie the observed reductions in tissue integrity within the amygdala, as demonstrated by studies reporting lower T1/T2 ratios59 and decreased grey matter concentrations60 in the amygdala of OSA patients, both reflecting the structural consequences of sustained ischemic and hypoxic stress in OSA. Additionally, one study reported elevated T2-relaxation values in the amygdala of anxious OSA patients, which may reflect increased free water content61. This increase could result from acute intermittent hypoxia causing vasogenic edema, or from chronic hypoxia causing axonal injury, demyelination, and gliosis. While these established findings would support a model in which OSA initiates structural changes that underpin the discriminatory texture features observed in our study, our Spearman correlation analysis revealed only subtle trends in this region. This lack of robust statistical significance suggests that the amygdala may be more resilient to respiratory-driven structural changes compared to the OFC in this pediatric population. However, the observed trends with sleep oxygen nadirs suggest that structural alterations, if present, are likely subtle and may only become prominent in cases of extreme nocturnal hypoxia. This clinical distinction may explain why emotional hyper-reactivity, an amygdala driven trait26, is often observed as secondary to executive deficits, an OFC driven trait15, in the clinical progression in the pediatric sleep-disordered breathing. Hippocampus The hippocampus is an important brain region for memory consolidation and learning28, and its dysfunction has been consistently linked to cognitive impairments observed in children with OSA62. To better understand the potential mechanisms underlying the dysfunction and structural changes, several studies have investigated how intermittent hypoxia in OSA affects hippocampal structure and function, particularly in relation to neurogenesis, cellular integrity, and inflammation. Liu et al.25 highlighted the highly active role of neurogenesis in the hippocampus. The review describes how prolonged intermittent hypoxia disrupts this process and ultimately impairs hippocampal development. Given that children are undergoing critical brain maturation, the suppression of neurogenesis in this region may contribute directly to the cognitive dysfunction observed in children with OSA. Additionally, inflammation may also play an important role in the dysfunction of the hippocampus. Inflammatory processes can contribute to neuronal injury through mechanisms such as oxidative stress, blood-brain barrier disruption, and microglial activation62. These combined effects of impaired neurogenesis and inflammation may help explain the hippocampal dysfunction seen in children with OSA. These disruptions can manifest as structural and metabolic abnormalities, which are detectable using advanced neuroimaging techniques. An MR spectroscopy study found that the hippocampus in children with OSA showed signals indicative of neuro-axonal dysfunction and active demyelination which are reflected in changes to specific brain metabolites40. This suggests damage to neurons, axons, and dendrites that may occur as a result of the disrupted environment caused by OSA. Microstructural changes have also been observed using diffusion tensor imaging (DTI), reporting reduced mean diffusivity (MD) values in the bilateral hippocampus of children with OSA38. MD is sensitive to changes in water diffusion across tissue barriers such as cell membranes and axonal structures. Decreased MD typically reflects acute cellular swelling, indicating that hippocampal tissue in these children may be undergoing early-stage injury. These findings highlight the structural and functional disruptions that are likely reflected in our significant texture analysis results. Notably, the Spearman correlation analysis demonstrated that hippocampal texture features are significantly associated with wake oxygen levels rather than exclusively with nocturnal parameters. This may suggest that systemic chronic hypoxia may have a lingering structural impact on memory-related regions, extending beyond the immediate sleep period. Clinically, these texture changes may represent early biomarkers of hippocampal atrophy or remodeling, providing a potential structural basis for the short-term memory and learning deficits, frequently observed in chronic pediatric sleep apnea patients independent of BMI39. OFC The OFC is a key region involved in executive functioning30, which is a domain that is often disrupted in children with OSA15. In the context of OSA, several studies have identified structural and functional impairments in the prefrontal cortex, which is a region that includes the OFC as an important subregion. More recently, Mingming et al. (2024) used resting-state functional near-infrared spectroscopy (rs-fNIRS) to assess prefrontal connectivity in adults with OSA63. The study found significantly reduced functional connectivity within the prefrontal cortex. These disruptions were associated with lower cognitive performance scores, thereby providing functional evidence of impaired network integration in the OFC and surrounding prefrontal areas. The authors also reported decreased global and local efficiency in the brain’s functional network and a reduction in connection edge numbers in key prefrontal circuits, which indicate weakened information processing and coordination within the OFC. To further contextualize these findings, Beebe and Gozal (2002) proposed a conceptual model linking the key pathophysiological features of OSA to impaired prefrontal cortex function through mechanisms of disrupted brain restoration and cellular stress64. Our Spearman correlation provides a neurobiological extension to this model, as higher OAHI and oxygen saturations correlated significantly with altered texture intensity and increased heterogeneity within the OFC. This suggests that the intermittent hypoxia and sleep fragmentation may be the primary drivers of structural micro-changes in this region. Such structural alterations could perhaps provide a biological basis for the deficits in decision-making, disinhibition, and emotional instability frequently reported in severe pediatric OSA. These findings highlight the vulnerability of the OFC and its direct contribution to the executive dysfunction commonly seen in children and adults with OSA. CPAP therapy CPAP therapy is the gold-standard treatment for Moderate/Severe OSA after adenotonsillectomy (AT) or when AT is not indicated. Its role in maintaining airway patency has been well established, improving oxygenation and reducing hypopnea events65,66. Previous studies have demonstrated that CPAP therapy in children with OSA can lead to significant improvements in cognitive function and overall quality of life67,68. Our study included 2 subjects with Moderate/Severe OSA that underwent CPAP therapy and received follow-up MRI scans. Our findings show that both CPAP-treated patients originally diagnosed with moderate-severe OSA now classify as not having OSA during their follow-up scans based on texture features in the ROIs. This can potentially suggest the importance of early intervention with CPAP in obese children with OSA to prevent or reverse subtle brain changes. Although the data of the two follow-up subjects with CPAP therapy were processed cross-sectionally, identical parameters were used for all time points to maintain consistency. Future studies with larger longitudinal samples should perhaps utilize a dedicated optimized longitudinal preprocessing pipeline to further minimize within subject variation. This would help to ensure the stability of these structural markers over time as well as the clinical utility of texture analysis in monitoring therapeutic response. Strengths of using the texture analysis feature selection and classification pipeline A critical step in the texture analysis pipeline is addressing the high dimensionality of extracted texture features to ensure accurate and interpretable results. Feature selection methods are particularly effective for identifying the most discriminative features while reducing noise and redundancy. Compared to other methods, the Fisher Score is a computationally efficient and fast feature selection method optimized for high dimensionality datasets. It attempts to achieve a reduced subset of statistical features that maximize the inter-class variance while minimizing intra-class variance within the feature data space69. The Fisher Score as part of a feature selection algorithm is commonly used in other clinical sleep apnea and systemic brain disease applications42,70–73. Once the most relevant features are selected, classification is applied to evaluate their significance. This is achieved by measuring the reduced dataset’s classification accuracy through a machine learning model, such as the k-nearest-neighbors (k-NN) model. k-NN is an instance-based, supervised machine learning algorithm, commonly used for diagnosis and disease prediction74. It is extremely adaptive, robust against noisy data, has a simple design, and is widely used for binary classification tasks75. k-NN commonly achieves the highest classification accuracy when compared to similar machine learning techniques43 and is frequently implemented in other systemic brain disease applications such as Parkinson’s disease, multiple sclerosis, and Alzheimer’s disease70,76,77. The algorithm is particularly well-suited for this application due to its simplicity and ability to handle complex, nonlinear relationships in high-dimensional data for small datasets78. This combination of advanced feature selection and classification offers a robust framework for identifying subtle brain alterations associated with OSA in obese pediatric subjects. Nevertheless, it is important to recognize that both methods have inherent limitations. Fisher analysis does not account for feature interactions or redundancies, which could limit its effectiveness in more complex feature spaces. Likewise, k-NN can be sensitive to the choice of the k parameter, as well as to variations in feature scaling and imbalanced datasets, potentially affecting performance in less controlled datasets. Although the class distribution between the groups in our dataset was relatively balanced, the overall sample size was small, which may limit the statistical power and generalizability of our findings. Smaller datasets always come with the risk of overfitting, especially with the high-dimensional feature sets derived from texture analysis are used. While the Fisher criterion helped reduce dimensionality, there is still potential for spurious associations or feature rankings that may not hold in larger or independent cohorts. This limited sample size also restricts the ability to explore potential confounding factors such as age, sex, or disease severity, and may mask subtle differences between the Moderate/Severe OSA and the non-OSA groups. While the current work utilized a focused contrast between Moderate/Severe OSA and No OSA, incorporating a broader spectrum of disease severity represents a logical next step. Future studies including a mild OSA cohort (OAHI 1.5–4.9 events/hour) would allow for a more nuanced observation of OSA-related brain changes across the full disease continuum, identifying the specific points at which imaging markers begin to deviate from neurotypical patterns. Moving toward these more inclusive, larger-scale datasets can help ensuring that selected texture features represent cohort-specific variance rather than a reflection of high-dimensional feature sets in a smaller cohort. Ultimately, this highlights the benefit for larger, multi-center studies for external validation, further establishing the clinical generalizability of our model. Furthermore, the absence of baseline neurocognitive or behavior testing prevents assessment of a definitive relationship between the observed texture features and functional outcomes. Consequently, these imaging findings should be interpreted as potential structural markers of injury rather than direct evidence of cognitive impairment. Further prospective research integrating standardized neuropsychological testing with texture analysis is essential to clarify the functional impact of these structural changes. Although T1-weighted data or advanced sequences such as diffusion tensor imaging (DTI) may offer important insights into tissue microstructure79,80, this study focused on T2-weighted data due to its optimal balance of pathological sensitivity and clinical utility. T2-weighted sequences are standard in routine clinical imaging and are uniquely capable of evaluating gray and white matter changes by assessing free water content, which increases in response to pathological processes such as vasogenic edema, gliosis, and demyelination following hypoxia-ischemia61. By leveraging texture analysis on high resolution T2-weighted images, our aim was to extract subtle structural information from routinely acquired sequences in clinical practice potentially enabling broader clinical translation and integration without requiring additional image post processing pipelines needed for DTI or other quantitative MRI sequences. While future multi-modal studies combining T1, T2, and diffusion metrics may further enhance sensitivity, our results demonstrate that texture features derived from standard T2-weighted images alone can potentially serve as accessible biomarkers for OSA-related brain changes. Clinical significance and conclusion The clinical significance of these findings lies in the potential of 3D texture analysis to act as a sensitive imaging tool for detecting subtle brain injury in obese children with OSA that may otherwise not be visible on conventional MRI through visual assessment. The amygdala, hippocampus, and OFC are regions that are closely linked to emotional regulation, memory, and executive function, respectively, all of which are commonly impaired in children with OSA. Our findings revealed microstructural alterations in these areas, consistent with prior evidence of OSA-related neuronal injury, disrupted neurogenesis, and inflammation. Early identification of these changes is crucial as these cognitive and behavioral deficits related to OSA can negatively affect well-being and development in children. Current clinical imaging lacks the sensitivity to detect these early and subtle changes, potentially delaying intervention. By detecting microstructural abnormalities in clinically relevant regions, texture analysis may guide timely diagnosis, risk stratification, and monitoring of treatment response in obese children with OSA. Therefore, integrating texture analysis into clinical workflows could improve early recognition of children at risk of neurocognitive dysfunction and support more personalized care. Further research is needed to validate these findings in a larger group integrated with neuropsychological testing in order to establish the value of texture features as a prognostic marker and determine how these markers may change in response to treatment such as CPAP. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Author contributions A.K and I.N were responsible for overseeing study completion and designing the experimental and neuroimaging acquisitions. S.T was responsible for ROI segmentation, literature review, and writing the first and subsequent drafts of the manuscript. D.S contributed to image analysis, development of figures, statistical analysis, writing methodology, and manuscript structuring and editing. A.R performed feature classification and machine learning aspects of the pipeline and development of ROC and AUC figures. A.K, I.N conceived the idea and design of the study. A.K provided the lead in manuscript preparation, including structuring, revising, and editing. I.N reviewed and edited the manuscript. Funding This work was supported by the Heart and Stroke Foundation of Canada [6250100011]. Data availability Data sharing is strictly controlled by our REB. The datasets from our study are available from the corresponding author on reasonable request and if granted permission by the REB. Declarations Competing interests The authors declare no competing interests. Ethics declaration Subject recruitment, anthropometric measurements, clinical assessments, and MRI acquisition were all performed at the Hospital for Sick Children in Toronto in accordance with all relevant guidelines and regulations. Informed consent was obtained from all individual participants included in the study and/or their legal guardian(s). All study protocols were approved by the institutional Research Ethics Board (REB Number: 1000050597). Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement Data sharing is strictly controlled by our REB. The datasets from our study are available from the corresponding author on reasonable request and if granted permission by the REB.