The current study enrolled children aged 1.5–6 years. A total of 50 participants aged 1.5–6.07 years (mean age: 3.12 ± 1.28 years) were included in the analysis, all recruited from Foshan Fosun Chancheng Hospital, China, between November 2021 and May 2023. All participants were assessed using the Gesell Developmental Schedules (GDS), developmental domains including fine motor, gross motor, personal-social, language, and adaptive behavior (23). Their parents or guardians completed the Autism Behavior Checklist (ABC) (24), a standardized questionnaire for assessing autistic behaviors and symptom severity. Children in the ASD group (n = 28; mean age: 3.24 ± 1.26; age range: 1.67–6.07 years) met the DSM-V criteria for ASD following a clinical interview and were assessed using the Autism Diagnostic Observation Schedule (ADOS Module 1 or 2) (25), administered by the same clinician. In addition, all children with ASD had ABC scores ≥ 53. Children in the TD group (n = 22; mean age: 2.97 ± 1.31 years; age range: 1.5–5.81 years) had GDS total scores > 86 and ABC scores < 47, indicating normal development and the absence of ASD. All children were native Mandarin and Cantonese speakers with normal hearing and without family history of mental or psychiatric disorders, including no family history of autism spectrum disorder.

Written informed consent was obtained from the parents or legal guardians of all participants in accordance with the Declaration of Helsinki. This study was approved by the Institutional Review Board of Foshan Chancheng Hospital (No. IRB-ATT-002-61).

To ensure that participants remained asleep during the MRI scan, oral chloral hydrate was administered for sedation at a dose of 0.5 mL/kg, with a maximum dose of 10 mL (at a 0.5% concentration) before scanning. All sedated children were continuously monitored by trained medical staff throughout the scan to ensure the safety of the participants.

DTI data were acquired from all participants using a 3.0 T SIEMENS Skyra scanner at Foshan Fosun Chancheng Hospital, China, with the following parameters: TR = 6800 ms, TE = 92 ms, slice thickness = 2 mm, matrix size = 128 × 128 mm, FOV = 220 × 220 mm², flip angle = 90°, gap = 0.2 mm, voxel size = 1.7 × 1.7 × 2 mm³, 60 directions, b-value = 1000 s/mm², 12 b0 image (b = 0 s/mm²), 56 slices, 72 volumes. The total scan duration was approximately 15 minutes.

Prior to data analysis, all DTI scans were manually inspected for artifacts and overall image quality. Because all scans were acquired under sedation, motion-related artifacts were minimal, and no scans were excluded based on quality concerns. The diffusion-weighted imaging data were processed using the Python Automated Fiber Quantification (pyAFQ; https://yeatmanlab.github.io/pyAFQ/; 16), an open-source tool that facilitates customized analysis. The “ParticipantAFQ” function was used to handle the entire workflow, including tractography, registration, segmentation, cleaning, profiling, and visualization. Preprocessing steps such as motion correction, eddy current correction, and tensor estimation were performed within “ParticipantAFQ” using the Vistasoft software package (https://github.com/vistalab/vistasoft). The tractography procedure used 25,000 seed points to generate streamlines through major white matter tracts. DTI metrics, including FA, MD, RD, and AD, were calculated for each tract. FA reflects the directionality of water diffusion and is often associated with fiber integrity; MD measures the overall magnitude of diffusion, indicating tissue density; RD reflects diffusion perpendicular to the main fiber direction, often used to assess myelination; and AD captures diffusion along the main fiber direction, providing insight into axonal integrity.

Tract diffusion profiles were extracted using the “ParticipantAFQ.export(‘profiles’)” command, generating structured arrays of tensor-based metrics for 28 tracts which were segmented into 100 equidistant nodes per group. Our analysis focused on four DTI metrics—FA, MD, AD, and RD—across 100 nodes in eight language-related tracts: the left and right AF, ILF, SLF, and IFOF.

The LI for each DTI metric was calculated using the formula: (R – L)/(R + L). A positive LI reflects greater values in the right hemisphere relative to the left, whereas a negative LI reflects greater values in the left hemisphere. One-sample t-tests were conducted in each group (ASD and TD) to assess whether the mean LI for each tract differed significantly from zero. Multiple-comparison correction was applied using 5,000-permutation testing (p < 0.05). For FA, which generally reflects microstructural integrity, significant positive t-values (corrected p < 0.05) indicate rightward lateralization, and significant negative t-values indicate leftward lateralization. For MD, RD, and AD, although higher values reflect greater diffusivity rather than greater integrity, the interpretation of the LI itself remains consistent: positive LI values indicate higher diffusivity in the right hemisphere, and negative values indicate higher diffusivity in the left hemisphere. Accordingly, significant negative t-values for these diffusivity metrics correspond to rightward lateralization, whereas significant positive t-values correspond to leftward lateralization.

To compare the LI between groups for each DTI metric and tract, liner regression models controlling for age and sex were conducted. Next, we conducted a regression analysis between the LI and language scores within each group using a linear regression model: LI ~ language scores + age + sex, controlling for age and sex. To evaluate whether the associations between LI and language scores differed between groups, we conducted a regression analysis that included the interaction term between group and language scores: LI ~ group * language scores + age + sex. Statistical significance was determined using a permutation-based multiple comparison correction with 5,000 permutations (p < 0.05).

We compared each DTI metric between the TD and ASD groups at both the tract-wise (mean value across 100 nodes) and point-wise levels using a linear regression model: DTI metric ~ group + age + sex, where group was the independent variable, age and sex as covariates. Statistical significance at the tract-wise level was determined using the permutation-based multiple comparison correction with 5,000 permutations (p < 0.05). For the point-wise analysis, consistent with previous research (26), only significant group differences (p < 0.05, permutation-based correction) observed across more than three consecutive nodes were considered significant.

We examined the relationships between each DTI metric and the GDS language scores separately for the ASD and TD groups at both the tract-wise (mean value across 100 nodes) and point-wise levels (mean value across more than three adjacent nodes showing significant group differences). The relationships were analyzed using the linear regression model: DTI metric ~ language scores + age + sex, controlling for age and sex. To evaluate whether the strength of these associations differed between groups, we conducted a regression analysis that included the interaction term between group and language scores: DTI metric ~ group * language scores + age + sex. Multiple comparisons were corrected using permutation testing with 5,000 permutations (p < 0.05).

To investigate the relationship between DTI metrics and symptom severity of ASD, we conducted a regression analysis between each DTI metric and total ADOS scores within the ASD group. This analysis was performed at both the tract-wise and point-wise levels (mean value across more than three adjacent nodes showing significant group differences) using a similar linear regression model: DTI metric ~ ADOS total scores + age + sex, controlling for age and sex. Permutation tests were applied to correct for multiple comparisons with 5,000 permutations (p < 0.05).

Participants in the TD and ASD groups were matched on age and sex (ps > 0.05). Children with ASD exhibited significantly higher ABC scores than TD peers (t₍₄₈₎ = -13.93, p < 0.001). However, children with ASD had lower GDS scores in five domains and total scores than TD children (ps < 0.001). Detailed demographic, clinical, and behavioral characteristics of both groups are provided in Table 1.

As shown in Table 2; Figure 1, there were significant lateralization patterns across DTI metrics after permutation-based correction. For FA, both TD and ASD groups showed significant rightward lateralization in the SLF (TD: t₍₂₁₎ = 2.34, p = 0.029; ASD: t₍₂₇₎ = 3.22, p = 0.003). For MD, significant lateralization was observed in the AF and SLF for both groups (all ps < 0.05); additionally, the ASD group showed rightward significant lateralization in the ILF (t₍₂₇₎ = -2.67, p = 0.013), whereas the TD group showed significant rightward lateralization in the IFOF (t₍₂₁₎ = -2.49, p = 0.021). For RD, significant lateralization was found in the IFOF and SLF for the TD group, and in the ILF and SLF for the ASD group (ps < 0.05). For AD, the TD group showed significant rightward lateralization in the ILF (t₍₂₁₎ = -2.55, p = 0.018), whereas the ASD group showed significant rightward lateralization in the AF (t₍₂₇₎ = -2.97, p = 0.006). However, the between-group lateralization analyses using linear regression models did not show any significant differences between the TD and ASD groups after permutation-based corrections (Supplementary Table S1).

As shown in Figure 2, a significantly positive association between the LI of FA and language scores was observed in the AF (β = 0.0008, SE = 0.0004, t₍₂₄₎ = 2.2, p = 0.037) in the ASD group, and a significantly positive associations between the LI of FA in the ILF and language scores (β = 0.002, SE = 0.0009, t₍₁₈₎ = 2.59, p = 0.019) was found in the TD group. There was also a significant group × language interaction for the LI of FA in the ILF (β = 0.002, SE = 0.0008, t₍₄₄₎ = 2.34, p = 0.02). No other significant associations were observed in the TD or ASD group after correcting for multiple comparisons.

At the tract-wise level, no significant group differences were found for any of the DTI metrics (Supplementary Figure S1). However, significant differences emerged at the point-wise level (Figure 3), where more than three consecutive significant nodes (p < 0.05, permutation corrected) were observed, following the approach used in previous research (26). No additional corrections for multiple comparisons were performed. Notable findings included significantly higher FA values in the ASD group compared to TD controls in specific regions: left AF (nodes 38–42), left SLF (nodes 82–90), left IFOF (nodes 53–56), right SLF (nodes 3–9, 66–72), and right IFOF (nodes 51–56). The ASD group showed reduced AD values in the left SLF (nodes 22–30) and right AF (nodes 17–20), but higher AD values in the right ILF (nodes 89–95) and right SLF (nodes 2–7, 66–69) compared to the TD group.

The ASD group demonstrated significantly negative associations between mean FA and language scores in the left AF (β = -0.001, SE = 0.0004, t₍₂₄₎ = -2.48, p = 0.021; Figure 4A), left ILF (β = -0.0008, SE = 0.0003, t₍₂₄₎ = -2.62, p = 0.015; Figure 4B), and right ILF (β = -0.0009, SE = 0.0003, t₍₂₄₎ = -3.59, p = 0.002; Figure 4C), controlling for age and sex. Additionally, the ASD group showed a significant positive association between mean RD in the left AF and language scores (β = 1.46e-06, SE = 5.6e-07, t₍₂₄₎ = 2.61, p = 0.015; Figure 4D). Following permutation-based correction, all other DTI metrics showed no significant association in either the ASD or TD group, nor any significant group × language interactions (all ps > 0.05).

Among associations between mean DTI metrics at nodes with significant point-wise group differences and language scores across various white matter tracts, the only significant finding was a positive association between mean AD in the right AF (nodes 17–20) and language scores in the ASD group (β = 2.64e-06, SE = 9.7e-07, t₍₂₄₎ = 2.73, p = 0.012; Figure 4E). No additional node-level associations survived permutation-based correction in either the ASD or TD group. Moreover, no significant group × language scores interaction was observed (p > 0.05).

Significant associations between the mean DTI metrics and ADOS total scores were observed in the ASD group after correcting for multiple comparisons (Figure 5). Significantly positive associations between mean FA and ADOS total scores were observed in the left AF (β = 0.003, SE = 0.001, t₍₂₄₎ = 2.43, p = 0.023) and a negative association between mean RD and ADOS total scores was noted in the left ILF (β = -4.7e-06, SE = 2.02e-06, t₍₂₄₎ = -2.3, p = 0.028). No other significant associations were found between mean AD and ADOS total scores in any tract after correcting for multiple comparisons using the permutation-based corrections.

Additionally, a significantly positive association was found between ADOS total scores and mean FA in the left AF (nodes 38–42) (β = 0.005, SE = 0.002, t₍₂₄₎ = 2.56, p = 0.017), and conversely, a significantly negative association was observed between ADOS total scores and mean AD in the right AF (nodes 17–20) (β = -8.08e-06, SE = 3.1e-06, t₍₂₄₎ = -2.61, p = 0.015).

Several limitations should be considered when interpreting these findings. Firstly, although permutation testing was applied to correct for multiple comparisons, the point-wise analyses remain exploratory and rely on relatively liberal statistical thresholds. As such, the point-wise results should be interpreted with caution. Secondly, while our point-wise analyses revealed detailed white matter changes in ASD, localizing the exact anatomical coordinates of these affected nodes remains challenging. This limitation, common in AFQ studies (38–40), arises from the reliance on fiber segmentation in individual native space. Thirdly, AFQ only analyzes the central portion of each fiber tract by clipping streamlines to the segment between the two defining ROIs, thereby focusing solely on the deep white matter “stem” of the tract (17). As a result, peripheral or branching segments are not evaluated, which may limit the ability to capture the full spatial extent of tract-specific alterations. Fourth, the DTI model cannot resolve crossing or branching fibers within a single voxel, which may lead to artificially reduced FA and altered MD, AD, or RD values in regions with complex white-matter architecture. Consequently, tract-based metrics—particularly in intersecting pathways such as the SLF, AF, and ILF—should be interpreted with caution. Finally, ASD symptoms may emerge later in development, particularly in children as young as those included in our control group. Future studies should incorporate longitudinal follow-up to determine whether any control participants later received an ASD diagnosis.