Dataset

The dataset used in this study consists of 2,940 facial images of children, collected from a publicly available Kaggle repository [30]. This corpus contains facial photographs of children with and without autism and is distributed for research use under the terms defined on its Kaggle page; we used the data accordingly and did not collect new images. The supervised task is binary screening (ASD vs. neurotypical) from a single RGB facial image. Sample images from this dataset, including both neurotypical and ASD-labeled cases, are illustrated in Figure 2 to provide visual context for the model’s input data. Class labels are provided by the dataset curators; no independent clinical re-verification was performed by the authors. This dataset is not a prospectively collected clinical cohort; labels typically come from the original public sources (e.g., caregiver/self-report) and therefore should be interpreted as screening-level, not as a confirmed medical diagnosis.

Display full size

Figure 2. Several examples of images of healthy (neurotypical) children and those with autism spectrum disorder [30]. Images are shown solely to illustrate input characteristics; identities are not disclosed, and presentation follows journal policies.

Prior studies using this Kaggle dataset report children approximately 2–14 years old (most cases in the 2–8 range), with images predominantly from Europe and the United States; therefore, the dataset is not demographically representative of global populations [31–33]. Exact age is not available for every child, and race/ethnicity metadata is generally missing, so age-stratified or fairness-by-ethnicity analysis is not possible.

Published reports note that many images were downloaded from autism-related websites and Facebook pages; labels are provided as binary ASD vs. non-ASD by the dataset curators [32]. We did not assign or modify these labels, and we did not collect any new images. In line with prior reports, the corpus comprises ~2,936 2D RGB images, and studies describe a near 1:1 class balance (autistic vs. non-autistic) [31, 32]. All images are standard RGB photographs; no infrared, depth, or multispectral imaging is available. The “multi-scale” term later in the paper refers to multi-scale feature extraction in the model, not to multispectral sensors. Because images originate from heterogeneous public web sources, the dataset exhibits uncontrolled pose, illumination, background clutter, occlusions/accessories (e.g., glasses, hats), and expression differences. This variability motivates our ROI-centric preprocessing, augmentation policy, and multi-scale modeling [32].

We performed an 80/20 stratified split (training/testing), preserving the ASD:non-ASD ratio and retaining an independent hold-out test set. Within the 80% training portion, we used five-fold cross-validation for validation/model selection (~64% train, ~16% validation, 20% test per fold). To prevent identity leakage, splits were performed at the subject level wherever multiple images per child were present. Augmentations were applied to training folds only; validation/test were never augmented.

We used a public, de-identified dataset and did not collect new data. For deployment, we advocate on-device inference, minimal data retention, and optional de-identification (e.g., landmark-only or blur-based strategies) to mitigate re-identification risks. Any eventual clinical or school use would require explicit guardian consent and formal approval by medical/ethical boards; this work does not claim clinical certification.

Prior work has shown that both facial data and other medical imaging data can expose sensitive identity information, and that de-identification strategies are needed to suppress identifiable cues while preserving clinically relevant content [34–37]. Our focus on local/on-device inference and ROI-level processing is aligned with this direction, because it avoids unnecessary external transmission and storage of full-face images.

Notably, the dataset used in this work consists of publicly available facial images of children that have been released for research use under the terms specified by the source repository. All analyses in this study were conducted on that existing dataset; no new images were collected. Because facial images of minors are highly sensitive, our framework is explicitly designed to minimize external exposure of the data: Inference is performed locally on the device after edge optimization, without uploading images to external servers. The model operates only on the cropped facial ROI, and we consider optional de-identification strategies such as facial landmark-based representations or controlled blurring for future deployment scenarios. We also note that any real-world screening application would require appropriate parental/guardian permission, explicit communication of purpose and data handling, and compliance with institutional and regulatory standards for working with minors.

Preprocessing and data augmentation

The data preparation process consists of three key steps: normalization, data augmentation, and ROI extraction. These steps enhance both the quality and diversity of the input data, contributing to improved model accuracy and efficiency. For each image, we first detect and crop a consistent facial ROI that centers on the child’s face and excludes most of the background. The ROI is aligned to include key facial structures such as the eyes, nose, mouth, and overall facial contour. This improves robustness to clutter, pose changes, and lighting variation in the original, in-the-wild images, and ensures that the classifier focuses on the child’s face rather than unrelated context.

In this work, when we refer to “craniofacial features”, we mean the overall geometric relationships and proportions of visible facial regions (for example, relative eye spacing, facial width-to-height ratio, nose-mouth distance, and periocular shape). These patterns have been discussed in prior autism-related facial studies as potentially discriminative cues. Here, we do not hand-engineer explicit measurements; instead, we provide the normalized face ROI to the model so that the MS-ViT can learn such cues directly from the image.

Normalization, a fundamental preprocessing step, standardizes pixel values to a consistent range (typically between 0 and 1), ensuring uniformity and precision in analysis. This process mitigates variations caused by lighting, contrast, and image quality, thereby enhancing the model’s feature extraction capabilities.

Figure 3 illustrates various data augmentation techniques applied to children’s facial images, each designed to enhance the diversity and robustness of the training dataset. These transformations—including rotation, horizontal flipping, cropping, brightness enhancement, contrast adjustment, noise injection, and blurring—simulate real-world variability in image acquisition, such as differences in lighting, orientation, and facial pose. By exposing the model to such diverse visual patterns, these augmentations help mitigate overfitting and significantly improve the model’s ability to generalize across unseen data. Augmentations are applied only to the training split; validation and test images are not augmented.

Display full size

Figure 3. Various data augmentation methods are applied to a facial image, including rotation, flipping, cropping, brightness and contrast adjustments, noise addition, and blurring. These steps improve dataset diversity and help the model generalize more effectively to real-world conditions.

Such preprocessing and augmentation steps are critical in real-time screening scenarios, where image quality and consistency cannot always be guaranteed. In the proposed MS-ViT and Edge AI hybrid framework, these techniques contribute directly to the accurate extraction of relevant facial cues associated with ASD risk from RGB images. By enriching the input space and stabilizing the facial ROI, the system becomes more resilient to pose, lighting, and occlusion. Any future use in clinical or educational settings would still require formal expert validation and appropriate ethical approval.

Model

The core of the proposed framework is MS-ViT, a multi-scale vision transformer. In our design, the same cropped facial ROI—an aligned RGB face crop of the child—is processed in parallel at multiple spatial resolutions (“scales”). Each scale-specific branch produces its own feature representation, and these representations are fused through a learnable weighted combination:

Here, X is the cropped RGB facial ROI; MS-ViT_i(X) is the feature map produced by the MS-ViT branch operating at spatial scale i; w_i is a learnable scalar weight controlling the contribution of scale i; n is the number of spatial scales (i.e., the number of parallel branches); and F_final is the fused multi-scale feature representation used for classification. The weights w_i are trained jointly with the rest of the network so that the model can automatically emphasize whichever spatial resolution is most informative. All inputs in this work are standard RGB face crops; no multispectral or infrared modalities are used.

To support low latency, on-device inference on resource-constrained hardware (e.g., Raspberry Pi-class devices), we apply a post-training edge optimization stage. This “edge optimization” step performs uniform quantization and structured channel pruning on the trained MS-ViT model. The goal is to reduce memory footprint and inference time without retraining the model from scratch or changing the task (binary ASD vs. neurotypical classification). Uniform b-bit linear quantization of a weight tensor W is defined as:

where W_q is the quantized integer tensor (values in [0, 2^b – 1]), W is the original full-precision weight tensor, b is the bit width (e.g., b = 8 for INT8), and min(W) and max(W) are the minimum and maximum values of W.

We further apply structured channel pruning to remove low-importance channels and reduce multiply-accumulate cost. Pruning can be expressed as an element-wise masking operation:

where W_p is the pruned weight tensor, M is a binary mask of the same shape (1 = keep, 0 = prune), and ⊙ denotes element-wise multiplication. We use structured (channel-level) pruning so that the resulting tensors remain efficient for integer execution on embedded CPUs. After pruning, the quantized/pruned model is exported in INT8 form and run locally on the device, avoiding the need to transmit children’s facial images to an external server.

All images in this work are standard RGB facial photographs of children. The term “multi-scale” refers to multiple spatial resolutions within the MS-ViT, not to multispectral sensing. After fusion, the final representation F_final is passed through fully connected layers for binary classification.

In the final decision-making phase, the extracted and fused features from multiple levels are processed through fully connected layers to perform the final classification. The supervised task is binary ASD screening: given a single face ROI (a child’s RGB facial image), the model outputs ASD vs. neurotypical. The learnable fusion weights allow the model to combine low-level cues (texture, local symmetry, contour sharpness) with higher-level craniofacial layout (relative positioning of eyes, nose, and mouth). This structure leverages multi-scale feature extraction as well as edge optimization to enable accurate predictions with low latency on resource-limited hardware.

This combination suggests technical feasibility for rapid, on-device ASD risk screening in settings such as schools or primary care clinics where computational resources and specialist access may be limited. Any real clinical or educational deployment, however, would require formal validation by medical experts, regulatory approval, and explicit guardian consent. Unlike a plain single-scale ViT, the proposed MS-ViT is co-designed with its edge optimization stage: it performs multi-scale facial feature fusion while also being compressible via quantization and pruning for local, privacy-preserving inference.

Training and hyperparameter tuning

The system is trained for supervised binary classification (ASD vs. neurotypical) from a single 224 × 224 RGB face ROI. We optimize a standard binary cross-entropy loss with L2 regularization:

Here, y_i ∈ {0, 1} is the ground-truth label for sample i (1 = ASD, 0 = neurotypical); p_i is the predicted ASD probability for that sample; N is the minibatch size; θ denotes all trainable model parameters; and $\lambda ‖\theta ‖\begin{array}{c}2\\ 2\end{array}$ is an L2 (weight decay) regularization term. Model selection is performed using mean validation AUC-ROC across cross-validation folds, and the final decision threshold is chosen on validation data and then fixed for evaluation on the held-out test set.

Hyperparameters (learning rate, weight decay, batch size, regularization/dropout, and augmentation magnitudes) are tuned via a compact iterative search over reasonable ranges, guided by validation AUC-ROC and training stability. We report threshold-agnostic metrics (AUC-ROC) directly. For thresholded metrics (accuracy, sensitivity, specificity), a single operating point τ^* is fixed from validation by maximizing Youden’s J :

and the same τ^* is used on the test set to avoid optimistic bias. All random seeds for splitting, augmentation, and optimization are fixed to support reproducibility. Calibration samples used later for PTQ are drawn exclusively from the training split and are described in the Edge-optimization subsection; the test set is never used for tuning, calibration, or threshold selection.

We performed a focused hyperparameter search to select the final training configuration. The parameters we tuned were the initial learning rate, weight decay, batch size, dropout strength, and augmentation strength (rotation, crop, brightness/contrast jitter). The learning rate was explored in the range 1 × 10^–5 to 1 × 10^–3; weight decay was varied between 1 × 10^–6 and 1 × 10^–3; batch size was varied between 16 and 64; and dropout probabilities were tested in the range 0.1–0.5. Data augmentation settings were varied within mild to moderate ranges, for example, in-plane rotation up to ± 15°, random cropping up to 10%, and brightness/contrast jitter up to ± 20%. For each candidate setting, we trained the model with the same supervised objective (binary cross-entropy with L2 regularization) and measured mean validation AUC-ROC across the five-fold cross-validation splits, using subject-level separation. Fewer than 30 unique configurations were evaluated. The best-performing configuration under this criterion was then retrained on the full 80% training portion and evaluated once on the held-out 20% test split, with the decision threshold τ^* fixed from validation as described in Equation (5).

Edge computing

To support real-time screening on embedded hardware without altering the trained decision function, we add a post-training edge optimization stage consisting of 8-bit quantization (INT8) and structured channel pruning. This stage is applied after full-precision training; validation and test partitions remain untouched and subject-level separation is preserved. The goal is to reduce latency, memory, and compute while keeping the model’s behavior aligned with the false positive 32 (FP32) network. Quantitative effects on runtime and accuracy are reported in Results. We adopt linear quantization for weights and activations with lightweight calibration drawn only from the training split (to avoid leakage). Dynamic range is estimated to compute the scale and zero-point,

and quantization/dequantization follow:

Weights use symmetric per-channel quantization in linear/convolutional blocks, while activations use asymmetric per-tensor quantization; normalization layers are folded where applicable. The exported INT8 graph runs integer kernels (ARM NEON) and is functionally equivalent to the FP32 network up to quantization noise; measured deltas are reported in Results. To further lower multiply-accumulate operations with predictable wall-clock gains on CPUs, we prune channels using an L1-norm importance score:

under a conservative global sparsity budget. A brief fine-tuning phase on the training folds restores any loss before quantization. We prefer structured (channel-level) over unstructured sparsity to preserve dense, kernel-friendly tensors and stable runtime on embedded devices. The pruned model is then quantized using the same PTQ procedure.

Experimental setup

To ensure reliable evaluation, the dataset was first split into 80% training and 20% testing subsets. The 80% training portion was then subjected to five-fold cross-validation, whereby the training data in each fold was further divided into new training and validation sets. Importantly, data augmentation techniques—including rotation, flipping, cropping, and contrast adjustment—were applied only to the new training subset in each fold, ensuring that validation and testing sets remained unaltered to avoid data leakage or overfitting.

Model training was conducted using the MS-ViT architecture, configured to capture multi-scale facial features. Input images were resized to 224 × 224 pixels and normalized to a 0–1 pixel range. The model was trained for 50 epochs using the Adam optimizer with a learning rate of 0.0001 and a batch size of 32. Early stopping with a patience of 10 epochs was used to prevent overfitting. Hyperparameter selection was based on grid search within the initial folds, and sensitivity analysis confirmed the model’s stability against minor variations in learning rate and batch size. On average, training each fold took approximately 18 minutes on an NVIDIA RTX 3080 GPU, while inference on edge devices (Raspberry Pi 4 with 4GB RAM) remained under 200 ms per image.

To further enhance real-time performance and deployability, the trained model was optimized using Edge AI techniques. Quantization (8-bit precision) and structured pruning were applied post-training to reduce the model’s size and computational load. These optimizations not only preserved model accuracy but also significantly improved inference speed in low-resource environments. The combination of rich, augmented data and an optimized, lightweight model provides a robust, accurate, and scalable solution for early autism detection using facial images.

Assessments

The evaluation of the proposed model was conducted using standard performance metrics, including accuracy, sensitivity, specificity, AUC-ROC, and inference time. These metrics confirm that the enhanced MS-ViT-based framework, equipped with Edge AI optimization and improved data augmentation, delivers superior performance in detecting autism from children’s facial images. Preprocessing and augmentation techniques were refined to increase dataset diversity and improve generalizability. The results of this evaluation are summarized in Table 1, comparing the fine-tuned model with its baseline and ablated variants.

In Table 1, the fine-tuned version of the proposed model (MS-ViT + Edge + Augmented) achieved an accuracy of 96.85%, with a sensitivity of 96.09% and a specificity of 97.92%, outperforming its baseline and reduced-feature versions. The AUC-ROC of 0.9874 indicates excellent class separation. Despite architectural improvements, the inference time remained efficient at ~181 ms per image, maintaining the model’s suitability for real-time deployment in resource-constrained environments.

This granular ablation cleanly separates optimization from architecture. The FP32 baseline reaches 96.12% at ~312 ms, whereas the deployed Edge variant (pruning + PTQ) attains 96.85% at ~181 ms. As expected, INT8 PTQ alone produces a negligible accuracy change (96.00%) with a sizable speed-up, while light structured pruning plus a brief fine-tune behaves as a regularizer on this relatively small dataset (96.55%). An optional QAT setup yields a modest improvement over FP32 (96.70%) at the same on-device latency, but still trails the combined pruning + PTQ path. Architectural ablations corroborate the contribution of each component—augmentation, ROI-centric preprocessing, multi-scale fusion, and sufficient backbone depth—each removal depresses performance despite broadly similar INT8-dominated latencies (~169–183 ms). Taken together, these results explain why “edge optimization” does not hurt accuracy here; if anything, the pruning stage slightly improves generalization, and the PTQ stage delivers the expected runtime gains.

Notably, consistent with Table 1 and Figure 4, the fine-tuned MS-ViT + Edge + Augmented model achieves 96.85% accuracy with 96.09% sensitivity and 97.92% specificity (AUC-ROC = 0.9874), while sustaining real-time throughput of ~181 ms/image on Raspberry Pi-class hardware. Across four independent runs, test accuracies cluster tightly around the table estimate, with balanced confusion matrices and low error rates, indicating stable behavior on ASD and neurotypical cases alike. These findings validate the importance of hierarchical multi-scale design and lightweight edge deployment, and support the framework’s suitability for screening in resource-constrained environments where both accuracy and latency matter.

Display full size

Figure 4. Four confusion matrices illustrating the performance of the fine-tuned MS-ViT model across different evaluation runs. Each matrix represents the classification outcomes on the test set, showing consistent high accuracy and minimal variation in prediction quality. ASD: autism spectrum disorder; MS-ViT: multi-scale vision transformer.

Figure 5 illustrates the comparative ROC curve analysis of the fine-tuned MS-ViT model and its baseline variant under two different evaluation scenarios. In scenario 1 (left), where the test images are drawn from standard lighting and frontal poses, the fine-tuned model achieves a superior AUC of 0.9865, clearly outperforming the baseline model, which achieves an AUC of 0.9796. This result demonstrates the effectiveness of the multi-scale feature extraction and fine-tuning in capturing subtle facial cues associated with autism. In scenario 2 (right), the test data simulates more challenging conditions, including slight variations in lighting and facial orientation. Although performance slightly decreases in this scenario, the fine-tuned model maintains a robust AUC of 0.9735 compared to 0.9590 for the baseline. These results not only confirm the model’s high discriminative capability across varying environments, but also highlight its generalizability and resilience—key attributes for real-world deployment in uncontrolled settings such as clinics or schools. The consistency in true positive rates and the sharp rise in ROC curves further indicate that the model preserves high sensitivity without compromising specificity, even under more demanding testing conditions.

Display full size

Figure 5. Comparison of ROC curves for the fine-tuned and baseline MS-ViT models under two evaluation scenarios. Scenario 1 represents standard testing conditions, while scenario 2 includes more challenging variations. The fine-tuned model consistently outperforms the baseline, achieving higher AUC values in both cases. AUC: area under the curve; ROC: receiver operating characteristic; MS-ViT: multi-scale vision transformer.

Conclusion

Throughout this study, we set out not just to improve technical performance in autism detection, but to make that progress meaningful in real-world terms—accessible, fast, and reliable. By combining the strengths of MS-ViTs with the efficiency of edge computing, our approach offers a practical solution for early autism screening using only facial images. The system achieved strong results, with high accuracy, sensitivity, and speed, even in low-resource settings. But beyond the numbers, the true value of this work lies in its potential to make early diagnosis more available to children who might otherwise be missed—especially in schools, rural clinics, or underserved communities. At the same time, we recognize that no single model can capture the full complexity of human development. Our method has limitations: it relies on still images, and its training data, while robust, doesn’t yet reflect the full diversity of the global population. These are not flaws—they’re frontiers. In future work, we aim to broaden our datasets, incorporate behavioral and temporal features, and refine the system for easier real-world integration. Ultimately, this research moves us one step closer to a world where early autism detection is not just possible, but practical—and where technology quietly supports the human work of care, understanding, and inclusion.