Prostate cancer is characterized by the uncontrolled growth of prostatic cells, driven by genetic, hormonal, and environmental factors. It represents the second most frequently diagnosed cancer in men, with an estimated 1.4 million cases and 375,000 deaths worldwide in 2020. Current detection methods include the prostate-specific antigen (PSA) test, digital rectal examination, and biopsy image analysis. Among these, biopsy analysis is particularly relevant, as it not only improves diagnostic accuracy but also guides treatment planning. Nevertheless, manual identification of malignant regions remains challenging due to the complexity of histopathological images. Consequently, deep learning approaches have gained attention as potential solutions. In this study, conventional architectures such as U-Net and FCN were employed alongside enhanced variants incorporating depthwise separable convolution (DSC). This modification substantially reduced the number of parameters and improved computational efficiency while maintaining, and in some cases surpassing, the accuracy of cancer tissue segmentation. Experimental results demonstrated that DSC-enhanced architectures significantly decreased runtime compared to baseline models. Furthermore, an ordinal loss function was integrated to address semantic segmentation across different Gleason scores, thereby increasing model sensitivity to the inherent ordinal structure of diagnostic categories.