Detecting cancerous anomalies in prostate MRI plays a vital role in the earlydiagnosis of prostate cancer. However, accurate identification remains highlychallenging due to the small size and heterogeneous morphology of cancer-ous regions, as well as the scarcity of fine-grained lesion annotations. To ad-dress these issues, we propose a Pseudo-Anomaly Generation and Frequency-Enhanced Network (PAF-Net) for prostate cancer detection. Specifically, wefirst train an ensemble of autoencoders and constrain their latent featurespaces to encourage diverse and redundancy-reduced representations of nor-mal anatomy. Then, guided by the reconstruction residual discrepanciesamong different autoencoders, we introduce residual-guided perturbationsin anatomically plausible regions to synthesize pseudo-anomaly samples as hard non-normal examples, thereby enriching the decision boundary withoutrequiring manual lesion annotations. Finally, we adopt a hybrid ConvNeXt–Swin detector as the backbone and further incorporate a frequency-domaincontrastive triplet learning strategy on the transformed features to enhancethe discriminability between normal and pseudo-anomalous patterns. Ex-perimental results on the PICAI, Prostate MRI-US, and Brain Tumor MRIdatasets demonstrate that our method achieves AUCs of 85.63%, 90.73%,and 90.83%, and APs of 82.74%, 88.26%, and 88.80%, respectively, outper-forming existing state-of-the-art methods and validating its effectiveness androbustness in prostate and brain MRI anomaly detection tasks.