Parkinson's disease is associated with the progressive death of dopaminergic neurons producing motor disorders at advanced stages. Today, there is no definitive biomarker to early detect and diagnose the disease (80% of the dopaminergic neurons are lost at first diagnosis). Recently, neuroimages have allowed to discover radiological markers associated with structural alterations from routine T1-MRI, and even measure dopamine activations in specialized SPECT sequences. Nonetheless, SPECT sequences are highly specialized, with low availability in clinical centers, and the respective analysis and quantification remain subjective and poorly explored. This work introduces a self-supervised deep representation able to capture dopaminergic levels from an MRI-to-SPECT transformation, learning an embedding representation with the capability to classify between Parkinson's and control subjects. From a retrospective study with a set of 73 Parkinson's and 73 control patients, the proposed approach achieved a notable discrimination of 79% in the precision metric.Clinical relevance- A self-supervised representation that under a translation MRI-to-SPECT task, can code dopaminergic levels to discriminate between control and Parkinson's subjects.