Parkinson's disease (PD) involves pathological iron accumulation, yet MRI metrics, such as R2* or magnetic susceptibility (χ), lack mechanistic specificity because they convolve paramagnetic and diamagnetic sources. We applied an AI-assisted χ-separation framework that combines deep learning (DL)-based preprocessing with biophysical modeling to assess paramagnetic iron with enhanced specificity. Twenty-five PD patients and twenty-six matched controls underwent 3 T multi-parametric MRI. DL-based χ-separation (χ-separationDL) separated the paramagnetic susceptibility component (χpara; indicative of iron) from χ, revealing alterations undetected by established susceptibility-based methods: χpara increased in dorsal premotor cortex ( +6.3%, P = 0.032) and substantia nigra pars compacta ( +10.2%, P = 0.024), χpara in premotor cortex correlated with disease duration (r = 0.41; P = 0.045). DL-based preprocessing was not inferior for the differentiation between PD patients vs. controls compared to established optimization-based χ-separation, indicating the potential for AI-enhanced χ-separation to be applied within the scope of susceptibility imaging in PD.