XAI-Enabled Equivariant Vision Transformer for Pediatric Pneumonia Detection

Abstract

Pneumonia remains the leading infectious cause of death in children under five, claiming over one million lives annually, with diagnostic delays in low-resource settings often exacerbated by rotated, flipped, or poorly aligned chest X-rays acquired from restless infants. Although convolutional neural networks (CNNs) and standard Vision Transformers (ViTs) have driven automated detection accuracies beyond 95% on benchmark datasets, their performance degrades significantly in real-world pediatric imaging due to limited geometric invariance and lack of clinical interpretability. This survey reviews advances reported recent years in deep learning approaches for pediatric pneumonia detection from chest X-rays, covering CNN-based hi- erarchies, global-context Vision Transformers, multimodal fusion with clinical markers, emerging equivariant transformer designs, and explainable AI techniques. Despite substantial progress in accuracy and sensitivity, persistent challenges include orienta- tion sensitivity, reliance on heavy data augmentation or non- imaging inputs, post-hoc interpretability, and limited standalone deployment in resource-constrained environments. The analy- sis highlights the need for geometrically robust, intrinsically interpretable, and clinically deployable models to bridge the gap between benchmark performance and reliable real-world pediatric screening.