Nutrient deficiencies affect over 2 billion people globally, with vulnerable populations disproportionately impacted. While artificial intelligence offers promising solutions for nutrient analysis, the lack of transparency in deep learning models limits their adoption in healthcare settings where interpretability is essential. This study presents a comparative analysis of interpretable deep learning models for nutrient analysis in food images, focusing on the trade-offs between model performance, interpretability, and computational efficiency. Using the Food101 dataset augmented with nutritional information, we implement and evaluate three distinct approaches: a ResNet50 architecture with Grad-CAM, a Vision Transformer with attention maps, and a DenseNet with LIME integration. Our hybrid approach, combining DenseNet with LIME, achieves 94% accuracy in nutrient detection whilst providing comprehensive interpretability, though at increased computational cost. The ResNet50 model demonstrates efficient performance with 92% accuracy and minimal computational overhead, whilst the Vision Transformer offers unique insights through attention mechanisms despite lower accuracy at 89%. Experimental results show varying effectiveness across nutrient types, with macronutrient detection achieving 95% accuracy and mineral detection proved more challenging at 87%. These findings provide valuable insights for selecting appropriate architecture based on specific application requirements, particularly where the balance between accuracy, interpretability, and computational efficiency is crucial.