Prostate cancer is an increasingly deadly global health concern, with current early screening and diagnostic methods facing severe limitations in performance and cost-effectiveness, leading to high rates of misdiagnosis and unnecessary biopsies. Gas chromatography-mass spectrometry (GC-MS) has been explored in the past with some success but lists of names of volatile organic compounds (VOCs) and their concentrations have failed to generalize across datasets and therefore we here present an alternative approach mimicking olfaction, extracting "scent character" signatures directly from GC-MS chromatograms of urine headspaces, without relying on molecular identification -a method inspired by medical detection trained dogs. Our methodology incorporates a comprehensive debiasing pipeline, including empirical Bayes correction, baseline drift removal, and source-bias adaptation through domain adversarial learning. GC-MS time series data are transformed into image representations suitable for convolutional neural networks, enabling the recognition and classification of prostate cancer scent features using a tool originally designed for image processing. The proposed model demonstrates classification performance consistent with canines in distinguishing cancer-positive from negative cases. Our findings highlight the potential of machine learning-driven scent characterization as a scalable and non-invasive Prostate cancer diagnostic tool, relevant to the wider emergent field of medical machine olfaction.