Analysis of acute stroke-like lesions in MELAS: Distribution, potential boundaries and spreading pattern Authors: Huada Tang, Junyu Liu, Xiying Cai, Fangda Leng, Jiaqi Hu, Yang Zhao, Zhaoxia Wang, Lei Yu Published: 2026 ABSTRACT BackgroundImaging during acute episodes of mitochondrial encephalomyopathy with lactic acidosis and stroke-like episodes (MELAS) primarily shows reversible cortical edema. However, the mechanisms underlying such cortical lesions in MELAS remain unclear. Analyzing MRI findings of MELAS could provide clues to these mechanisms.MethodsWe reviewed the medical records and imaging data of MELAS patients treated at Peking University First Hospital from 2013 to 2024. 61 MRI scans from 55 patients were included in this study. The spatial distribution of SLLs was analyzed based on high-resolution cortical segmentation. Next, we established a non-negative independent component analysis (nnICA) model to separate independent SLL components. Then we assessed correlation of synaptic density and lesion load using Spearman's rho test. Finally, a simple diffusion model was used to fit the spreading behavior of SLLs.ResultsThe cortical probability map shows a peak involvement frequency of 0.46 observed in the left temporo-occipital lobe. Distribution of SLLs showed no bias between left and right hemisphere. Then we identified 8 independent cortical components of SLLs. The Probability Map showed a positive correlation with synaptic density map (Spearman's rho = 0.32, p < 0.001). Additionally, the Directed-Effect Coefficient exhibited a good proportional relationship with the Dilation Coefficient (y = 0.50x+0.00, r = 0.76).ConclusionAnalysis of structural MRI data suggests that MELAS SLLs exhibit a pattern of free spreading within regions and restricted spreading between regions on the cortex. The identified components indicate the presence of potential boundaries. We also showed that lesion spreading was often hindered in areas with sparse synaptic density. Introduction Stroke-like episode (SLE) is the characteristic feature of Mitochondrial encephalomyopathy with lactic acidosis and stroke-like episodes (MELAS), commonly presenting as acute or subacute stroke-like symptoms such as headache, unilateral visual disturbances and focal-onset seizures(Tetsuka et al., 2021; Sproule et al., 2008). The mechanisms underlying SLEs have been a topic of debate for decades. Three main hypothetical mechanisms were the mitochondrial angiopathy (vascular hypothesis) theory, the mitochondrial cytopathy theory, and the neuronal hyperexcitability theory(Tetsuka et al., 2021; Iizuka et al., 2005). Recent evidence suggests that neuronal hyperexcitability may be the true initiating cause of MELAS(Iizuka and Sakai, 2009). MRI plays an irreplaceable role in the diagnosis and management of MELAS. In the acute phase of SLE, brain MRI in patients often shows stroke-like lesions (SLLs) confined to cortex and subcortical white matter(Sproule et al., 2008; Malhotra et al., 2016). Classic SLLs are characterized by a continuous distribution of lesions that does not correspond to the territory of major blood vessels(Alves et al., 2019). The MRI findings in MELAS patients change over time relative to the onset of episodes. In the acute phase, MRI often shows linear high signals on Diffusion-Weighted Imaging (DWI) and low signals on Apparent Diffusion Coefficient (ADC)(Cheng et al., 2022; Ito et al., 2008; Kim et al., 2011). T2 and T2FLAIR show cortical hyperintensity and edema, which gradually extend to the subcortical regions. At the same time, the oxygen extraction fraction (OEF) in the lesion area significantly decreases, indicating deterioration of mitochondrial function in the local brain tissue(Xie, 2014; Yu et al., 2013). Accumulation of metabolic waste and excitatory neurotransmitters, together with altered neuro-glial coupling leads to vasodilation in the lesion area, resulting in significantly increased local cerebral blood flow (CBF) perfusion (Li et al., 2017; Rodan et al., 2015). In the subacute phase, the high signals on DWI gradually fade, and ADC shifts to iso-intensity or high signal. The hyperintensities on T2 and T2FLAIR take longer to return to normal(Ito et al., 2008). The Classical SLLs are asymmetric, predominantly involving the posterior brain regions, including the temporoparietal junction, parietal lobe, and occipital lobe(Alves et al., 2019). Several studies have provided cortical distribution probability maps of SLLs, but they all lack sufficient sample sizes(Wang et al., 2020a, 2024). Additionally, some researchers have suggested that MELAS lesions exhibit a characteristic pattern of slow spreading(Wang et al., 2020a; Iizuka et al., 2003), but the factors influencing the spreading remain unknown. In signal processing, Independent Component Analysis (ICA) is a computational method for separating mixed information generated by multiple independent signal sources. Its classic application is the “cocktail party problem,” which involves isolating each person's voice in a noisy environment. Our study innovatively used ICA method to explore the characteristics of the spreading pattern of MELAS lesions. Triggers for MELAS stroke-like episode include cold exposure, overeating, physical exhaustion, and stress. These factors can disrupt local neuronal energy metabolism and blood flow, exceeding the threshold of self-regulation in MELAS patients. As a result, MELAS patients may develop a ‘lesion core’ on the cortex. Neurons in the affected area exhibit hyperexcitability due to insufficient inhibition and an unstable membrane potential(Iizuka et al., 2005, 2012). Subsequently, the hyperexcitability at the lesion core can influence closely connected electric and metabolic networks, leading to widespread neuronal dysfunction. When examining the lesions in MELAS patients, we can observe that many lesions exhibit remarkably similar distribution, which suggests the existence of potential boundaries that constrain the spread of lesions. It is natural to assume that the entire cortex can be divided into Independent Components, where acute lesions spread freely within a component region but are restricted at the boundaries between components. The generation of SLLs can be interpreted as: 1) Following the triggering factors, independent lesion centers emerge within several subregions of the patient's cortex. 2) These lesion centers spread freely within components but stop when they encounter components boundaries. 3) The SLLs on a single MRI image are actually a combination of spreading lesions derived from lesion centers mentioned above. Visually interpretation of this process is showed in Fig. 1. Following assumptions above, we conducted a detailed quantitative analysis of the spreading characteristics of SLLs.Fig. 1Schematic of Acute SLL Processa. A schematic diagram of one of the neuronal hyperexcitability mechanisms in MELAS patients. ATP depletion occurs due to a disruption in the mitochondrial energy supply chain. Then the Na+/K+-ATPase fails to maintain the polarized membrane potential. The extracellular high potassium levels can stimulate Na+/K+-ATPase activity, further depleting ATP and forming a positive feedback loop. As a result, the excitation threshold of neurons is reduced. b. Disturbances can lead to the emergence of independent lesion centers in the cortex, which are represented by red gradient point light sources. c. Schematic of the diffusion model for nnICA algorithm. The red and green ribbon-like regions represent different diffusion areas, corresponding to distinct independent components in ICA. Red and green gradient light sources represent lesions in two regions, the diffusion process of lesions is relatively free within regions and is hindered when crossing regions' boundary. The nnICA algorithm is to detect whether there exists such subregions and boundaries. Schematic of Acute SLL Process a. A schematic diagram of one of the neuronal hyperexcitability mechanisms in MELAS patients. ATP depletion occurs due to a disruption in the mitochondrial energy supply chain. Then the Na+/K+-ATPase fails to maintain the polarized membrane potential. The extracellular high potassium levels can stimulate Na+/K+-ATPase activity, further depleting ATP and forming a positive feedback loop. As a result, the excitation threshold of neurons is reduced. b. Disturbances can lead to the emergence of independent lesion centers in the cortex, which are represented by red gradient point light sources. c. Schematic of the diffusion model for nnICA algorithm. The red and green ribbon-like regions represent different diffusion areas, corresponding to distinct independent components in ICA. Red and green gradient light sources represent lesions in two regions, the diffusion process of lesions is relatively free within regions and is hindered when crossing regions' boundary. The nnICA algorithm is to detect whether there exists such subregions and boundaries. Materials and methods Data acquisition Ethical statement and data availability The requirement for informed consent was waived as this was a retrospective study using fully anonymized data with no direct interaction with participants. Anonymized data not included in the article due to space constraints can be made available upon request to any qualified investigator for the purpose of replicating the study's procedures and results. All procedures in this study were performed in accordance with relevant laws, institutional guidelines, and the Declaration of Helsinki. The study protocol was approved by the Ethics Committee of Peking University First Hospital on November 17, 2024 (approval number: 2024-279-002). Participants We reviewed the medical records of all patients who were clinically diagnosed as ‘MELAS’ at our center between 2013 and 2024. These patients were then screened according to the following diagnostic criteria: Category A: Clinical findings of stroke-like episodes, which include 1. Classical manifestation of mitochondrial stroke-like episodes(Ng et al., 2019) and 2. Acute stroke-like lesions determined by brain MRI. Category B: Evidence of mitochondrial dysfunction, which includes 1. High lactate levels in plasma and/or cerebrospinal fluid or deficiency of mitochondrial-related enzyme activities, 2. Mitochondrial abnormalities in muscle biopsy and 3. Definitive gene mutation related to MELAS. We included patients who simultaneously met two items of Category A and two items of Category B, which was identified as Definite MELAS in several studies(Yatsuga et al., 2012; Lorenzoni et al., 2015; Gonçalves et al., 2020). Patients with a history of other neurological disorders or suspected head trauma were excluded in our study. MRI acquisition MRI examinations were performed using a 3.0-T MR scanner (Signa Excite™) equipped with an 8-channel head coil. The routine imaging sequences included T1-weighted axial images (axial T1WI), T2-weighted axial images (axial T2WI), T2-fluid attenuated inversion recovery axial images (axial T2 FLAIR), diffusion-weighted axial images (axial DWI), and T1-weighted sagittal images (sagittal T1WI). The parameters for T1WI were: TR = 2340 ms, TE = 9.6 ms, FOV = 240 mm, slice thickness = 6.0 mm, and voxel size = 0.4688 mm × 0.4688 mm × 6.0 mm. For T2WI, the parameters were: TR = 5000 ms, TE = 110 ms, FOV = 240 mm, slice thickness = 6.0 mm, and voxel size = 0.4688 mm × 0.4688 mm × 6.0 mm. The T2 FLAIR parameters were: TR = 9602 ms, TE = 114.2 ms, slice thickness = 6.0 mm, and voxel size = 0.9583 mm × 0.9583 mm × 6.0 mm. The parameters for DWI were: TR = 4000 ms, TE = 62.4 ms, FOV = 240 mm, slice thickness = 6.0 mm, and voxel size = 0.9583 mm × 0.9583 mm × 6.0 mm, b = 1000 s/mm2. MRI data selection After patient inclusion, we reviewed all MRI scans for each individual and included those at acute stage: 1) DWI showed hyperintensity along the cortex. 2) There was a stroke-like episode nearby (within one month). Selected MRI examinations should contain at least T2FLAIR, T1/T2 and DWI sequences. If one patient undergone multiple MRI scans, all scans meeting the above criteria were included. Acute Lesion Mask In order to reflect characters of acute SLLs, Acute Lesion Mask (ALM) is defined as the cortical area affected by the most recent stroke-like episode. The interactive software ITK-SNAP was used to delineate ALMs on the patients' MRI images. Hyperintensities on DWI were used to identify the location of fresh lesions, then gyral swelling and hyperintensity on T2FLAIR image were manually segmented and proofread by three researchers with over 5 years of professional experience. Raw data preprocessing Image registration was processed in two steps using ITK-SNAP(Yushkevich et al., 2006) and SimpleITK. We used ICBM 2009c Nonlinear Symmetric template as the registration reference. The patients’ structural T1/T2 images were first compared to the standard template on ITK-SNAP and were roughly aligned manually. Then the resampled images were automatically normalized using B-spline transformation algorithm. Next, the ALMs, which represented acute SLLs, were transformed to standard Montreal Neurological Institute (MNI) space. Finally, the transformed ALMs were overlaid on standard template and transformed T1/T2 image to check the results. Fig. 2 illustrates the registration process described above.Fig. 2Raw Data Preprocessing Schematica. Registration flowchart, the final transformed image was compared with standard template to show registration quality. b. ALM transformation overlaid on T2 image. We used the insular region delineated by researchers on the raw image for demonstration. It can be observed that the transformed insula area aligns well with the insular region on the standard template. c. Schematic of NAV ALM vector.ALM: Acute Lesion Mask, NAV: Normalized Affected Volume. Raw Data Preprocessing Schematic a. Registration flowchart, the final transformed image was compared with standard template to show registration quality. b. ALM transformation overlaid on T2 image. We used the insular region delineated by researchers on the raw image for demonstration. It can be observed that the transformed insula area aligns well with the insular region on the standard template. c. Schematic of NAV ALM vector. ALM: Acute Lesion Mask, NAV: Normalized Affected Volume. In order to analyze the spatial distribution of acute SLLs, we introduced the CerebrA atlas to segment cortical regions. CerebrA is based on an accurate non-linear registration of cortical and subcortical labelling from Mindboggle 101 to the symmetric MNI-ICBM2009c atlas(Manera et al., 2020). To avoid confusion, we refer to the regions in the CerebrA atlas as CerebrA segments. By overlaying both ALMs and CerebrA segments on standard template, we can describe the location of the acute SLLs as following. For an ALM associated with a specific MRI scan, the Normalized Affected Volume (NAV) for a given CerebrA segment is defined as the number of voxels in the intersection of the transformed ALM and the CerebrA segment in MNI space. The Affected Coefficients (AC) is defined as the ratio of the NAV to the total number of voxels in this CerebrA segment. In this way, each ALM can be described as two vectors of size n (the number of cortical CerebrA segments), with elements representing NAV and AC respectively. Statistical analysis Calculating 95% confidence intervals To better show the involvement probability for each CerebrA segment, we used the bootstrap method to compute the 95% confidence interval of mean values. Bias in lesion distribution between left and right hemispheres We examined whether ALMs had similar distributions in symmetric CerebrA segments between the left and right hemispheres. For a given CerebrA segment X, the distribution was compared between NAV for X in left hemisphere and X in right hemisphere using Kolmogorov-Smirnov (K-S) test. Besides, we also applied Mann-Whitney-U test on the same data to examine whether the average NAV for left and right CerebrA segments were the same. A P value < 0.05 was considered statistically significant. Non-negative ICA model Non-negative ICA We used nnICA technique introduced by M. D. Plumbley(Plumbley et al., 2004), and we referred to the nnICA algorithm from Marius Langed's GitHub to process our data (https://github.com/Marius1311/Non-negative-ICA). The application of non-negative ICA requires the following prerequisites to be met: Firstly, the signal sources are independent from each other. From the discussion in the background part, the signal sources in our model are lesion cores in component regions, and they are essentially independent, given that the potential boundaries block the spreading behavior. In this way, signal strength of a specific component can be defined as the lesion size within this component's region. Approximately, overlapping size of lesion area derived from this signal and CerebrA segment X is proportional to signal strength. As a result, the NAV for CerebrA segment X is just the linear combination of each signal's contribution, which correspond to the second prerequisite of nnICA. The last condition is that the signal strengths are continuous non-negative variables, our data satisfied this condition immediately. Data matrix preprocessing The raw ALM data matrix derived from arranging all NAV ALM vectors into a matrix. We discarded those CerebrA segments never affected and those do not belong to the cortex. Given that the symmetric CerebrA segments in the left and right hemispheres have the same ALM distribution (Supplementary Table 2), we only considered the ALM distribution of a single hemisphere to further reduce data dimensionality and increase the data volume. If both hemispheres are simultaneously affected by SLLs, lesions in each hemisphere were separated and obtained two ALMs. If an ALM involved patient's left hemisphere, then the ALM was symmetrically mapped to right hemisphere (Fig. 5a). In this way every ALM was represented by CerebrA segments in right hemisphere. Finally, we can get 74 ALM vectors from 61 MRI scans, indicating that 13 scans involved simultaneous acute lesions in both hemispheres. Examples of truly bilateral involvement were presented in Supplementary Fig. 1.Each ALM vector's length is 28, corresponding to NAVs for 28 CerebrA segments in right hemisphere's cortex. To improve ICA component identification, we performed Principal Component Analysis (PCA) on the NAV data matrix. We selected a variance explanation ratio of 95% and extracted a total of 8 principal components. The data was then mapped into the principal component space, resulting in an 8×74 data matrix, which was the nnICA algorithm applied on. Probabilistic map's correlation with synaptic density The synaptic density map was the recent result from Knudsen(Johansen et al., 2024). They combined in vivo and in vitro method to obtain a 3D atlas of synaptic marker SV2A. We used a cortical mask to obtain the synaptic density distribution excluding the frontal cortex and calculated the Spearman's rho and p value with our SLL's distribution. Then we used a bootstrap method to obtain 95% confidence interval of the Spearman's rho value. To validate the true association, we controlled for potential confounding factors by incorporating cortical thickness and an FDG-PET signal template into the analysis. Specifically, cortical thickness was derived from the fsaverage standard template, while the FDG-PET template was generated as the average of 116 normal Chinese 18F-FDG PET brain images(Wang et al., 2021) (https://github.com/DlutMedimgGroup/Chinese-Brain-PET-Template). To further assess the independent contribution of synaptic density, a multiple linear regression model was fitted with lesion load as the dependent variable and SV2A, PET signal, and cortical thickness as predictors. All variables were standardized prior to analysis to facilitate comparison of effect sizes. Partial correlation analyses were performed to assess the association between synaptic density and lesion load while regressing out PET signal and cortical thickness. Spreading characteristic analysis Directed-Effect If acute lesions simultaneously involve both CerebrA segments X and Y, then we can say X and Y are co-involved. We defined the Directed-Effect Coefficient as a metric to measure the degree of co-involvement: recall that every ALM had AC for X and Y, denote as ACX and ACY. We can compute CORRXY = CORR (ACX, ACY) as the correlation coefficient of CerebrA segments X and Y, which measures the degree of linear correlation between lesions involvement within X and Y. Besides, we can fit a regression line between ACY and ACX, the slope of this line k reflects the extent to which the lesion in CerebrA segment X affects segment Y. In this way, Directed-Effect Coefficient from X to Y was defined as CORRXY×kX-Y. Dilation Coefficient We established a simple model to measure the degree of adjacency between two CerebrA segments on the cortex. Starting from one CerebrA segment, the first iteration was defined as: If a cortical voxel is in contact with at least one voxel from this CerebrA segment, then it was added to the region. The following iteration steps were similar. This iteration was designed to slightly expand the existing region along the cortex. The Dilation Coefficient from CerebrA segment X to Y was defined as: 1) Dilating outward from X for 65 iterations. 2) Each iteration has overlapping area with CerebrA segment Y, and the overlap ratio is defined as the proportion of the overlapping area relative to CerebrA segment Y. 3) Computing the average overlapping ratio for 1-65 iteration as the dilation coefficient. This process was visually described in Fig. 7d. Results Observational statistics Demographics and clinical data We included 55 patients in our study according to the diagnostic criteria, comprising 30 males and 25 females (P = 0.501). A total of 61 MRI scans of these patients were included. The patients' age was ranged from 9 to 50, with an average age of 23 (SD = 10.8) at the first included MRI examination. Table 1 summarizes the clinical characteristics of symptoms across each entire SLE event, including newly emerging symptoms during disease progression. Seizures were the most common manifestation (28/61, 46%), followed by headache (16/61, 26%), visual symptoms (14/61, 23%), and neuropsychiatric symptoms (12/61, 20%). We further analyzed the distribution of initial presenting symptoms at SLE onset, which showed a similar ranking (Table 1). Genetic testing was conducted on 51 patients, revealing that 47 patients (92%) had the typical m.3243A > G mutation, while the remaining patients were found to have one instance each of the mutations m.14453G > A(Yokota et al., 2020), m.14487T > C (Xu et al., 2019), m.14502T > C, and m.10158T > C (Wang et al., 2020b). Although m.14502T > C often occurs in Leber's hereditary optic neuropathy, this patient exhibited typical stroke-like episodes and characteristic MELAS MRI lesions, so he was diagnosed with ‘MELAS’ and included in this study.Table 1Clinical information of included patients and SLE events.Table 1FeaturesCounts (SDs, proportions)p valuesTotal patients55 male30 (54%)0.501 female25 (46%)Total SLEs61Age (years, of total SLEs)23 (10.8)MRI time from SLE onset15.9 (12.8)Symptoms in current SLE headache16 (26%) visual symptoms14 (23%) seizures28 (46%) numbness3 (5%) hemiplegia (limb weakness)12 (20%)/ auditory symptoms8 (13%) nausea and vomiting6 (10%) aphasia6 (10%) neuropsychiatric symptoms12 (20%)Onset symptoms headache12 (20%) visual symptoms8 (13%) seizures20 (33%) numbness2 (3%) hemiplegia (limb weakness)8 (13%) auditory symptoms4 (7%) nausea and vomiting1 (2%) aphasia5 (8%) neuropsychiatric symptoms1 (2%)This table summarized the clinical characteristics of the enrolled 55 patients and 61 SLE events. The sex distribution was balanced (54% male and 46% female). Patients' age and time from SLE onset to MRI were also calculated. We showed detailed symptoms' proportion for both initial manifestations and whole SLE events. Visual symptoms include blurred vision, visual disturbances and visual field defect. Table 1Clinical information of included patients and SLE events.Table 1FeaturesCounts (SDs, proportions)p valuesTotal patients55 male30 (54%)0.501 female25 (46%)Total SLEs61Age (years, of total SLEs)23 (10.8)MRI time from SLE onset15.9 (12.8)Symptoms in current SLE headache16 (26%) visual symptoms14 (23%) seizures28 (46%) numbness3 (5%) hemiplegia (limb weakness)12 (20%)/ auditory symptoms8 (13%) nausea and vomiting6 (10%) aphasia6 (10%) neuropsychiatric symptoms12 (20%)Onset symptoms headache12 (20%) visual symptoms8 (13%) seizures20 (33%) numbness2 (3%) hemiplegia (limb weakness)8 (13%) auditory symptoms4 (7%) nausea and vomiting1 (2%) aphasia5 (8%) neuropsychiatric symptoms1 (2%)This table summarized the clinical characteristics of the enrolled 55 patients and 61 SLE events. The sex distribution was balanced (54% male and 46% female). Patients' age and time from SLE onset to MRI were also calculated. We showed detailed symptoms' proportion for both initial manifestations and whole SLE events. Visual symptoms include blurred vision, visual disturbances and visual field defect. Clinical information of included patients and SLE events. This table summarized the clinical characteristics of the enrolled 55 patients and 61 SLE events. The sex distribution was balanced (54% male and 46% female). Patients' age and time from SLE onset to MRI were also calculated. We showed detailed symptoms' proportion for both initial manifestations and whole SLE events. Visual symptoms include blurred vision, visual disturbances and visual field defect. Acute SLL's spatial distribution The cortical probability map of CerebrA segments (Fig. 3) shows that acute SLL is primarily located in the temporal, occipital, and parietal lobes, with a peak involvement frequency of 0.46 observed in the left temporo-occipital lobe. This cortical heatmap was generated using the MATLAB plugin BrainNet Viewer(Xia et al., 2013). The average AC and NAV for all CerebrA segments (in standard MNI space) are presented in Supplementary Table 1. Additionally, Fig. 4 provides the names, bar plots, and scatter plots of the top 10 CerebrA segments with the highest average NAV and AC respectively, along with their spatial locations in standard MNI space.Fig. 3Probability Map of Acute SLLs Over the Cortex.Registered SLLs were overlaid and projected to the cortex. The colors in the color bar represent the frequency of involvement of the corresponding cortical regions.Fig. 4Top-10 CerebrA segments by involvement proportion or volume.a. Top-10 CerebrA segments by involvement proportion (mean AC) and their visualization over cortex.b. Top-10 CerebrA segments by involvement volume (mean NAV) and their visualization over cortex.Fig. 5Schematic of nnICA processa. If the acute SLL involved patient's left hemisphere, then the ALM was symmetrically mapped to right hemisphere. b. First ALM Matrix was reduced to 8 dimensions by PCA, and the nnICA algorithm was applied on ALM Matrix in PCA space. After undergoing the PCA inverse mapping operation, the Unit Source Matrix was obtained. c. Visualization of the Unit Source Matrix using heatmap. The color of each square in the heatmap represents the proportion of the component in the corresponding row occupying the CerebrA segment in the corresponding column. Probability Map of Acute SLLs Over the Cortex. Registered SLLs were overlaid and projected to the cortex. The colors in the color bar represent the frequency of involvement of the corresponding cortical regions. Top-10 CerebrA segments by involvement proportion or volume. a. Top-10 CerebrA segments by involvement proportion (mean AC) and their visualization over cortex. b. Top-10 CerebrA segments by involvement volume (mean NAV) and their visualization over cortex. Schematic of nnICA process a. If the acute SLL involved patient's left hemisphere, then the ALM was symmetrically mapped to right hemisphere. b. First ALM Matrix was reduced to 8 dimensions by PCA, and the nnICA algorithm was applied on ALM Matrix in PCA space. After undergoing the PCA inverse mapping operation, the Unit Source Matrix was obtained. c. Visualization of the Unit Source Matrix using heatmap. The color of each square in the heatmap represents the proportion of the component in the corresponding row occupying the CerebrA segment in the corresponding column. Bias in lesion distribution between left and right hemispheres Acute SLLs in MELAS rarely exhibit continuous cross-hemispheric distributions, which is probably associated with the minimal involvement of deep white matter in MELAS. However, whether the subregions of the left and right hemispheres exhibit similar lesion involvement patterns in MELAS has yet to be explored. We found that the average rates of acute SLL involvement were roughly equivalent between the left and right hemispheres. Only 3 CerebrA segments (Insula, Pars Opercularis, Precentral) showed significant difference (Supplementary Table 2). Moreover, except for Precentral, all CerebrA segments showed quite similar distribution of NAV between left and right hemispheres. This finding supports the approach of treating bilateral hemisphere involvement as two separate and independent datasets for processing and analysis if both hemispheres are affected in a patient. nnICA components As mentioned in the Methods section, after preprocessing, we obtained a 28×74 data matrix, where 28 represents the number of CerebrA segments and 74 corresponds to the number of ALMs within a single hemisphere. Following nnICA, the ALM matrix was decomposed into two matrices: the Unit Source Matrix and the ICA Weights Matrix in the independent components' space. Specifically, the Unit Source Matrix, sized 28×8, defines the independent components. Each unit source is characterized by its expected involvement in the 28 CerebrA segments, illustrating the cortical extent and strength of this component. On the other hand, the ICA Weights Matrix represents each ALM in the independent components' space, where each ALM is expressed as a combination of multiple components with varying weights. According to our discussion of nnICA background in Method section, an acute SLL was a combination of several independent components of different strength, these strengths, or weights are captured in ICA Weights Matrix. Fig. 5 provides a detailed and intuitive illustration of the aforementioned process. Fig. 6 is the visualization of 8 independent components. Each component was presented using a cortical heatmap along with sagittal, coronal, and axial cross-sectional views. The ordering of the components represented how well they are separated by the nnICA algorithm. The most dominant component was named Supra-Middle Temporal Zone, predominantly occupies CerebrA segments Middle Temporal, Superior Temporal and Transverse Temporal. The other 7 components were named as Occipital Zone, Frontal Zone, Post-Parietal Zone, Parieto-Temporal Junction Zone, Insula Zone, Inferior Temporal Zone and Medial Occipital Zone, respectively. For each independent component we selected one typical patient's image and exhibit in Supplementary Fig. 2.Fig. 6nnICA Components VisualizationEach component was showed by Cortical map and image slice from axial, coronal and sagittal respectively. The color bar represents an involvement possibility for each component of unit signal strength.Fig. 7Synaptic Density Effect and Neighborhood-Spreading Characteristic.a. Spearman test result of synaptic density map and SLLs distribution map. Each voxel was represented by a green circle in the scatter plot. b. Scatter plot and linear fitting of Directed-Effect Coefficients and Dilation Coefficients for every ordered pair (X, Y), where X and Y traversed all CerebrA segments except for the frontal lobe. c. Heatmap of Directed-Effect Coefficients. The order of the CerebrA segments has been adjusted to ensure that cortical regions that are nearby be arranged as close to each other as possible. d. Schematic of the dilation iteration model. Red cube in the center denotes voxel in pre-existing area. After one iteration, the adjacent cortical voxels (light blue) were included as part of the new region. nnICA Components Visualization Each component was showed by Cortical map and image slice from axial, coronal and sagittal respectively. The color bar represents an involvement possibility for each component of unit signal strength. Synaptic Density Effect and Neighborhood-Spreading Characteristic. a. Spearman test result of synaptic density map and SLLs distribution map. Each voxel was represented by a green circle in the scatter plot. b. Scatter plot and linear fitting of Directed-Effect Coefficients and Dilation Coefficients for every ordered pair (X, Y), where X and Y traversed all CerebrA segments except for the frontal lobe. c. Heatmap of Directed-Effect Coefficients. The order of the CerebrA segments has been adjusted to ensure that cortical regions that are nearby be arranged as close to each other as possible. d. Schematic of the dilation iteration model. Red cube in the center denotes voxel in pre-existing area. After one iteration, the adjacent cortical voxels (light blue) were included as part of the new region. The mirroring strategy is designed to reduce data dimensionality and increase data representation, facilitating the identification of less dominant independent components. To validate the rationale of this approach and the robustness of the identified components, we performed a sensitivity analysis without hemispheric mirroring. Specifically, nnICA algorithm was applied on unmirrored lesions’ data matrix to identify another set of independent components (unmirrored-ICs). The results showed good consistency: most ICs could be matched with corresponding component maps in unmirrored-ICs, except for IC5 and IC7. This supports the robustness of the cortical components of MELAS lesions identified in our analysis (Supplementary Figs. 3 and 4). Synaptic density and lesions’ boundaries We got rho = 0.325 and p < 0.001 after applying Spearman test on Synaptic density map and SLL distribution map, which showed a significant moderate positive correlation. We only considered the cortex outside the frontal lobe because frontal lobe lesions are very rare. The 95% confidence interval for the rho value is 0.320–0.328. The scatter plot of synaptic density value and lesion load for each voxel was showed in Fig. 7a. Furthermore, potential confounding factors were considered to test the robustness of the association between synaptic density and lesion load. Multivariate regression analysis showed that synaptic density had the strongest independent association with lesion distribution (0.301). Compared to synaptic density, PET signal demonstrated a weaker association with lesion load (beta = 0.13), while cortical thickness showed only a minimal contribution (beta = - 0.05) (Supplementary Fig. 6A). Supplementary Fig. 6B shows the partial correlation analysis after controlling for cortical thickness and PET signal. The residual correlation between synaptic density and lesion load remained largely preserved (r = 0.286), closely comparable to the unadjusted correlation (r = 0.325), supporting an independent association between synaptic density and lesion load. In order to better exhibit the tight correlations between synaptic density and SLL's boundary, some examples were showed on Supplementary Fig. 5. The termination of acute lesions often occurs in areas with synaptic density troughs and thinner cortical thickness. Spreading characteristic analysis For each CerebrA segments pair X and Y, the Directed-Effect Coefficient from X to Y and the Dilation Coefficient from X to Y have been defined in the method section. Fig. 7b is the visualization of correlation between Directed-Effect Coefficient and Dilation Coefficient. Each data point on the plot was an ordered CerebrA segments pair (X, Y). Due to the rarity of lesions in the frontal region and their high randomness, data points for frontal CerebrA segments were not displayed in the figure. The Directed-Effect Coefficient exhibited a good linear relationship with the Dilation Coefficient, with the fitted line given by y = 0.50x+0.00, r = 0.76. This suggests that the closer the cortical distance between sites A and B, the more likely the lesion at site A is to spread to site B. Then the Directed-Effect Coefficient for each ordered CerebrA segments pair was showed by the heatmap (Fig. 7c). We arranged the display order of CerebrA segments so that CerebrA segments belonging to the same major brain area (e.g., frontal, parietal, occipital, temporal) were positioned as closely as possible. Evidently, high-brightness region pairs are mostly located on either side of the diagonal. The heatmap reveals a trend: the closer the anatomical relationship, the higher the Directed-Effect Coefficient. Fig. 7b and c together demonstrate that cortical anatomical adjacency is the primary pattern of lesion's spreading in MELAS. Discussion This study represented the largest known structural MRI investigation of MELAS to date. We conducted an in-depth analysis of acute SLLs in MELAS patients, yielding a series of intriguing findings that hold significant value for elucidating the pathogenesis of MELAS. First, descriptive statistics revealed the spatial distribution of lesions, consistent with previous studies(Wang et al., 2020a, 2024; Bhatia et al., 2020), showing that acute lesions predominantly occur in the temporal, occipital and parietal regions. Comparative analysis of the left and right hemispheres demonstrated that the distribution of acute lesions exhibits remarkable bilateral symmetry. Many researchers have found that acute lesions in MELAS exhibit a gradual cortical spreading pattern(Iizuka et al., 2003). However, it is evident that the cortical spreading of acute lesions is not homogeneous but instead exhibits regionalization. For instance, lesions in the frontal lobe rarely spread to other cortical regions, and vice versa. To analyze the factors influencing lesion spreading, we innovatively applied a non-negative independent component analysis (nnICA) algorithm, which roughly divided the lesion signals into eight independent sources. Furthermore, nnICA was applied to the data matrix while preserving lesion laterality, yielding similar component results, supporting the robustness of the cortical components we identified. The greatest significance of these independent components is that they suggest the boundaries of the acute lesions' spreading in MELAS. As we mentioned earlier, the spreading of MELAS lesions is hindered when crossing boundaries, resulting in well-defined edema regions on MRI. Firstly, a clear boundary exists between frontal lesions and posterior lesions, defined by structures like the (Caudal Middle Frontal)-(Precentral) boundary and the Pars Opercularis-Insula boundary. Secondly, there exists clear boundary between Supra-Middle Temporal Zone and Occipital Zone, which is roughly located at the junction between the Inferior Temporal and Lateral Occipital CerebrA segments. These boundaries may possess different neural electrochemical connectivity structures from those within the components’ regions. Surprisingly, these potential boundaries show a high degree of overlap with regions of sparse synaptic density. Recently researchers used in vivo and in vitro methods to create a 3D atlas of synaptic density(Johansen et al., 2024). They observed the lowest density in the lateral occipital cortices and the sensorimotor cortices, which are boundaries of our independent components. Furthermore, the cortical synaptic density map has a moderate positive correlation to our SLL's distribution (Fig. 7), and the confounding factor analysis supports an independent association between synaptic density and lesion distribution. Lesion boundaries often appear at synaptic density troughs, reflecting the role of these synapse-sparse regions in blocking lesion spread. However, mechanism behind this fact need further exploration and the rarity of frontal lobe lesions cannot be explained in this way. Besides, quantitative susceptibility mapping (QSM) may offer complementary information about iron distribution and susceptibility alterations in MELAS lesions. Combined QSM/R2∗ approaches have been shown to distinguish iron- and myelin-related components in ischemic lesions, it may be useful for connecting structural lesion distribution with biochemical and microstructural pathology(Uchida et al., 2022, 2024). To quantitatively analyze whether MELAS lesions evolve through an adjacent-spreading pattern, we designed the Dilation Coefficient and Directed-Effect Coefficient and analyzed correlation between them. The results showed that the Dilation Coefficient and Directed-Effect Coefficient exhibit a positive proportional correlation, while high Directed-Effect rates often observed between CerebrA segments in close cortical proximity, reflecting a simple adjacent-spreading pattern. The present framework primarily captures the spatial (diffusion-like) component of lesion propagation. However, the biological process is likely governed by an interplay between local nonlinear dynamics. Future studies integrating reaction–diffusion models or connectome-based approaches may provide a more comprehensive description of these processes. Conclusions This study conducted an in-depth investigation into the distribution patterns of acute stroke-like lesions in MELAS patients, innovatively utilizing blind source separation techniques to identify eight primary independent components (lesion origins). Furthermore, quantitative analysis verified the adjacent-spread pattern of acute lesions, which aligned with our experience. Our results identified boundaries that hinder lesions spread, which shows an extraordinary match with synaptic density maps. These intriguing findings suggested that the cortical structures at the boundaries merit further exploration, which could help elucidate the nature of central nervous system lesions in MELAS. Limitations First, the relatively coarse slice thickness of most MRI scans may affect the accuracy of lesion identification and registration. But the registration is relatively accurate and has small error, as shown in Fig. 2. Second, we performed non-negative ICA decomposition on the NAV matrix rather than in MNI space. This resulted in our independent components being based on CerebrA segments rather than voxels. Besides, we extracted only eight independent components because PCA was applied for dimensionality reduction, which does not mean there are only eight independent diffusion regions. Because left-hemisphere lesions were mirrored to the right hemisphere to reduce dimensionality, subtle hemispheric asymmetries may have been attenuated. Our spreading analysis was based on a simplified diffusion-like framework, which cannot reproduce the nonlinear biological processes that likely underlie SLL propagation, such as excitotoxicity, metabolic collapse, and disrupted neurovascular coupling. Finally, although we have verified that synaptic density is closely related to both lesion involvement and lesion diffusion boundaries, this cannot be simply interpreted as synaptic connectivity density determines lesion spread. Their causal relationship requires further investigation and detailed explanation. Declarations and statement This study was supported by National High Level Hospital Clinical Research Funding (High Quality Clinical Research Project of Peking University First Hospital) No.2023HQ03. The authors have no relevant financial or non-financial interests to disclose. This study has been approved by the Ethics Committee of Peking University First Hospital. The requirement for informed consent was waived as this was a retrospective study using fully anonymized data with no direct interaction with participants. Anonymized data not included in the article due to space constraints can be made available upon request to any qualified investigator for the purpose of replicating the study's procedures and results. CRediT authorship contribution statement Huada Tang: Data curation, Formal analysis, Investigation, Methodology, Software, Visualization, Writing – original draft, Writing – review & editing. Junyu Liu: Data curation, Investigation, Software, Writing – original draft, Writing – review & editing. Xiying Cai: Data curation, Investigation, Software, Writing – original draft, Writing – review & editing. Fangda Leng: Conceptualization, Data curation, Methodology, Software, Writing – review & editing. Jiaqi Hu: Data curation, Investigation. Yang Zhao: Data curation, Methodology, Software. Zhaoxia Wang: Conceptualization, Data curation, Funding acquisition, Project administration, Resources, Supervision, Writing – review & editing. Lei Yu: Data curation, Funding acquisition, Supervision, Validation. Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.