Deep Bridge: A Unified Convolutional Neural Network Framework for Dual-Task Healthcare Applications in Autism Spectrum Disorder Screening and Sign Language Recognition
Jul 2026· 2026 4th International Conference on Sustainable Computing and Smart Systems (ICSCSS)· pp. 477-481· 0 citations· 23 references
Abstract
In this paper, we propose Deep Bridge, a unified Convolutional Neural Network (CNN) model that can be considered as a dual task model for healthcare, to solve two important problems: Autism Spectrum Disorder ( ASD ) screening from facial image analysis and recognizing American Sign Language ( ASL ) for better accessibility. It is a hierarchical Multi-Layer Perceptron (MLP) with 512-256-128-64 neurons in each layer, Rectified Linear Unit (ReLU) as an activation function, and Adam optimizer with early stopping regularization. Sixteen 64x64 RGB images were flattened to 12,288-dimensional feature vectors which are required for expression recognition, eye contact pattern and hand gestures essential for classification. Experimental evaluation on synthetic benchmark and clinical data demonstrates strong performance: 94.50% and 96.25% accuracy on autism screening and sign language recognition respectively, with weighted F1-scores of 0.945 and 0.962. In binary classification of autism, the Area Under the Receiver Operating Characteristic Curve (AUC-ROC) is 0.967, which indicates high discriminative power. Statistical robustness is validated through 5-fold cross-validation yielding mean accuracies of 93.82% (+/-1.24%) and 95.68% (+/-0.87%). A web-based Streamlit interface enables real-time screening, with an average of 23.4ms for inference, making it a clinically deployable solution that connects research and clinical application of machine learning.
Early non-invasive screening technologies are a paramount priority in modern healthcare for identifying complex neurodevelopmental traits characterized by social, communicative, and behavioural challenges. Recent breakthroughs in computer vision and deep learning have established automated facial image analysis as a highly viable paradigm for objective clinical screening. This study introduces a robust, optimized weighted ensemble framework that integrates the complementary architectural strengths of convolutional and transformer-based networks for binary classification of these specialized facial trait profiles. The pipeline concurrently leverages EfficientNet-B5 for localized feature scaling, Data-Efficient Image Transformers (DeiT) for long-range global self-attention, and ConvNeXt for modernized, high-performance convolutional representations. To ensure generalization and counteract dataset selection bias, a strict stratified 5-fold cross-validation scheme is enforced, followed by an optimized out-of-fold weighted probability fusion mechanism. Experimental evaluation on a benchmark dataset demonstrates that the unified ensemble achieves a state-of-the-art classification accuracy of 95.67% and an ROC-AUC of 0.9788, significantly outperforming individual standalone baselines. These empirical results validate that bridging high-frequency local textures with low-frequency global contextual relationships minimizes predictive variance, offering an accurate, stable, and scalable computational screening solution for automated clinical environments.
B. Anjali, S. Gopinathan· International Journal on Inf...· 0 citations
The proposed VGG16-based approach has potential as a supportive, non-invasive tool for early ASD screening and is deployed as an interactive, Streamlit-based web application that allows users to upload facial images and receive real-time predictions.
: Early identification and proper classification of brain tumours are essential to making the correct clinical decision. Magnetic Resonance Imaging (MRI) provides high resolution structural data, but the tumor variability in terms of size, shape and texture are a common limiting factor in the performance of traditional Computer-Aided Diagnosis (CAD) systems. In this study, an augmented radial basis function networks (ARBFN), novel long-short-term memory repeat neural networks (ILSTM-RNN), and efficient regularized convolutional neural networks along with dimension reduction modules (ERCNN-DRM) are proposed as a form of deep learning-based CAD framework. MRI images formatted in DICOM are subjected to bilateral filtering, semantic segmentation, and feature reduction through Principal Component Analysis (PCA) to maintain distinctive characteristics. The ERCNN-DRM classifier reached an accuracy of 95.7%, surpassing current techniques in both sensitivity and specificity, while considerably lowering computational demands. The results obtained in the experiments prove the possibility of the framework in the effective and automatic classification of benign and malignant brain tumours that precondition its use as the promising solution of intelligent computer-assisted healthcare systems.
Pallavi G. Patil, Anuja Chincholkar, N. Jagtap et al.· Proceedings of the 1st Inter...· 0 citations
These findings demonstrate the potential of computer vision-based analysis of children’s expressive activities as an effective, non-invasive ASD pre-screening tool and will focus on expanding dataset diversity and integrating multimodal behavioral cues to improve model generalization and clinical applicability.
Aina Khairina Ahmad Khair, Wan Mohd Yaakob Wan Bejuri, Mohd Murtadha Mohamad et al.· Bulletin of Electrical Engin...· 0 citations
Autism spectrum disorder (ASD) consists of a spectrum of neurodevelopmental conditions characterized by complex behavioural traits and subtle, atypical facial morphologies. Analysing these facial biomarkers provides a promising, non‐invasive avenue for objective clinical screening, addressing the subjectivity of traditional diagnostic processes. Therefore, the present study aims to introduce DHCA‐Net, a deep learning framework designed for automated ASD detection through the analysis of facial images. The proposed architecture uses different levels of attention and different adaptive tuning to advance discriminative learning. Specifically, DHCA‐Net introduces three novel attention mechanisms to advance discriminative learning: a hierarchical channel module for spatial‐semantic processing, a temporal‐depth convolutional module for local contextual control and an inverted residual multi‐core module for dynamic feature learning. An adaptive refinement step is also introduced to denoise clinical features. To bolster and diversify the classifier, the model exploits deep spatial and contextual resources via a multi‐head feature fusion (MHFF) mechanism. We conducted extensive testing on a benchmark dataset comprising 2936 diverse, preprocessed facial images (86.4% train, 10.2% test, 3.4% validation splits). A comparative analysis evaluated DHCA‐Net against architectures such as DenseNet, Xception, EfficientNet and Swin‐Transformer. The proposed model achieved 93.7% classification accuracy, a 0.9365 F1 score and a 0.9887 AUC, demonstrating superior performance. Furthermore, the model maintains an efficient average inference latency of 70.14 ms despite its mid‐to‐high computational complexity. This explainable framework offers significant clinical applicability for scalable screening, though future work must address generalization across broader demographic populations.
Davinder Paul Singh, Tathagat Banerjee, A. C et al.· International Journal of Dev...· 0 citations
Alzheimer's disease is a leading cause of death with no cure. Therefore, early detection is critical to slow progression and preserve quality of life. Diagnosis relies on medical history, cognitive tests, physical exams, and MRI brain scans, making deep learning suitable for Alzheimer's classification. This work proposes a benchmark that evaluates ten different convolutional neural network (CNN) architectures (including ResNet, DenseNet, MobileNet, EfficientNet, and VGG family models) under the same held-out test split protocol. A two-stage transfer learning and full fine-tuning pipeline is introduced to perform training using a class-balanced subset (3,900 images) derived from the OASIS medical imaging dataset, comprising 86,437 single-view MRI brain scans labeled into four classifications of Alzheimer's disease: Non-Demented, Very Mild Dementia, Mild Dementia, and Moderate Dementia. The best results were achieved by VGG16, with a 0.9637 validation accuracy and a 0.9533 test accuracy score. A key finding documented in this work is the difficulty of classifying the transition from Non-Demented to Very Mild Demented stages, observed consistently across all ten architectures.
