The early detection of nutritional deficiencies in crops is a key challenge in precision agriculture, as it directly impacts yield and crop quality. This study evaluates various object detection model architectures aimed at identifying nutritional deficiencies in lettuce plants (Lactuca sativa L.) through image analysis. Key performance metrics, including loss function, mean Average Precision (mAP), ROC curves, confusion matrices, precision, recall, and F1-score, were analyzed. Transformer-based models, such as DDQ and Conditional DETR, demonstrated superior performance, with mAP values exceeding 0.90 and high discriminative capability, albeit at the cost of increased computational complexity. In contrast, lighter alternatives like FoveaBox and PAA struck an optimal balance between efficiency and accuracy, making them suitable for deployment on resource-constrained devices. The results highlighted that high loss values do not necessarily indicate poor model performance, underscoring the importance of utilizing multiple evaluation metrics for a comprehensive assessment. These findings confirm the potential of deep learning in precision agriculture, facilitating the automated and accurate detection of nutritional deficiencies, and contributing to the optimization of agronomic management and decision-making in the field. Finally, recommendations for future research and practical applications are provided to encourage the adoption of these technologies in real-world production settings.