Accurately detecting music symbols in images of historical, complex, dense orchestral or piano printed scores can be challenging due to old printing techniques or time degradations. Because segmentation problems can vary widely, a data driven approach like the use of deep learning detectors is needed. However, the production of detection annotations (symbol bounding boxes + classes) for such systems is costly and time consuming. We propose to train such model with synthetic data and annotations produced by a music typesetting program. We analyze which classes are relevant to the detection task and present a first selection of music score typesetting files that will be used for training. To evaluate our model, we plan to compute quantitative results on a synthetic test set and provide qualitative results on a few manually annotated historical music scores.