Down syndrome (Trisomy-21) is a common congenital genetic condition associated with distinct dysmorphic facial phenotypes. Early screening plays a critical role in clinical management and developmental support. However, training deep convolutional neural network pipelines for genetic classification is heavily restricted by data scarcity and class imbalances inherent to medical imaging cohorts. This paper addresses these limitations by introducing a robust, multi-stage framework that couples a fine-tuned Deep Convolutional Generative Adversarial Network (DCGAN) with the EfficientNet architecture. To circumvent the high-resolution instabilities common in generative optimization, our architecture implements a clear multi-stage scaling hierarchy: synthesis is executed natively at \(64\times 64\) pixels utilizing a mathematically smooth Softplus activation combined with \(R_1\) gradient regularization to enforce local Nash equilibria, followed by an optimized bi-cubic transformation to upscale the augmented feature space to \(224\times 224\) pixels. Using a curated baseline dataset of 3,328 original frontal facial photographs, the generative pipeline synthesized 665,000 high-fidelity phenotypic examples, achieving a Fréchet Inception Distance (FID) score of 9.4951 and KID score of 7.6097(the KID metric has been scaled by multiplying by 1000). Ablation studies demonstrate that generative data expansion provides an approximate 12% absolute boost over standard deep learning methods. The proposed model framework achieves a definitive classification accuracy of 97.58%, precision of 97.19%, recall of 97.08%, and an F1-score of 97.13%. Interpretability validation via Gradient-weighted Class Activation Mapping (Grad-CAM) confirms that the classification head systematically extracts clinically relevant landmarks, suggesting the potential of the proposed framework as a non-invasive decision-support screening aid. The system is not intended to replace clinical assessment, genetic testing, or karyotype-based diagnosis.
