Pneumonia remains a leading cause of morbidity and mortality worldwide, particularly in low-resource settings where early diagnosis is critical. Chest X-rays are commonly used for detection, yet accurate segmentation of pneumonia-affected regions remains a challenge due to overlapping structures, vague boundaries, and variations in lesion appearance. Conventional loss functions often treat all pixels equally, limiting the model’s ability to focus on diagnostically relevant areas. This study proposes a novel loss function, Feature-Prioritized Dice Loss (FPD Loss), which integrates statistical feature importance and entropy-based region weighting to guide deep learning models toward clinically significant regions during training. The method is implemented using U-Net and DeepLabv3+ architectures and evaluated against standard Dice Loss and Cross-Entropy Loss. A comprehensive experimental setup includes augmentation strategies, early stopping, learning rate scheduling, and robustness testing under synthetic noise conditions. Quantitative results demonstrate that FPD Loss achieves superior segmentation performance across multiple metrics, including Dice Score, IoU, Precision, Recall, and Sensitivity. Furthermore, the proposed method exhibits increased resilience to perturbations, suggesting enhanced generalization. These findings underscore the effectiveness of feature-aware optimization in advancing automated pneumonia diagnosis from chest X-ray images.