Identification of Parkinson’s disease (PD) patients at risk for development of dementia is crucial for early intervention. However, diagnosing dementia in PD patients requires the use of a time-consuming and complex battery of psychological tests performed by an experienced psychologist. The study aims to prove the usefulness of convolutional neural networks for the identification of brain areas related to the progress of cognitive impairment by using standard magnetic resonance imaging (MRI) sequences. T1 & T2 sequences of 18 patients were used in the pilot study. Activation maps were generated, and the brain regions most involved in the classification process were identified, showing areas potentially significant in the diagnosis of cognitive impairment severity. The cerebellum was proven significant for distinguishing the above-mentioned classes in relative cerebellum volume (ANOVA p value = 0.0038 with large effect size = 0.5254) and folding (p value = 0.0031, = 0.5357), which is consistent with reports by other authors. Our analysis demonstrates that convolutional neural networks combined with a proper image preprocessing pipeline could be used for feature extraction in MRI sequences and can successfully support the identification of disease-specific abnormalities of the brain image.