Bladder, kidney, and prostate cancers are prevalent urinary cancers, and developing efficient detection methods is of significance for the early diagnosis of them. However, noninvasive and sensitive detection of urinary cancers still challenges traditional techniques. In this study, we developed a SERS-based method to analyze serum samples from patients with urinary cancers. Rapid, label-free, and highly sensitive detection of human sera is achieved by cleaning and aggregating silver nanoparticles. Furthermore, a long short-term memory deep learning algorithm is used to distinguish serum spectra, and the performance of the model is evaluated by comparing the accuracy, sensitivity, specificity, and receiver operating characteristic curves. Taking advantage of SERS and machine learning in sensitivity and data processing, the three urinary cancers are clearly classified. This is the first attempt to exploit the SERS-machine learning strategy to discriminate multiple urinary cancers with clinical serum samples, and our results showed the potential application of this method in the early diagnosis and screening of cancers.