This paper presents the DS4DH team’s approaches for the ImageCLEFmedical Caption 2025 challenge, where we participated in two tasksConcept Detection and Caption Prediction. For Concept Detection, unlike the typical approach of multi-label classification, we posed the problem as image to sequence of tokens mapping, where the tokens consist of the CUI’s, and a standard transformer maps the image embeddings into the sequence. Our approach achieved an F1-score of 52.3%, ranking top-6 among the best submission systems. For Caption Prediction, we developed multiple approaches, including fine-tuned InstructBLIP, traditional retrieval-augmented generation (RAG), and cluster-based RAG methods. Our best strategy, based on the InstructBLIP model, achieved the highest recall (BERTScore (Recall) of 60.7%) among all participants, being ranked top-2 according to the Overall challenge metric (33.6%). Our experiments reveal that RAG approaches did not outperform the baseline, exposing critical challenges in medical image captioning where noisy retrievals significantly weaken generation performance. Through validation experiments and case studies, we demonstrate that only highly accurate reference images prove helpful, as poor retrieval quality introduces noise that degrades caption generation.