Bladder cancer staging is necessary for the prognosis of patients and planning appropriate treatment. Computed Tomography-based tumor staging is still limited by inter-observer variability and differing diagnostic opinions. This work presents SliceNet, a deep learning approach to classifying bladder cancer as Stage II, III, and IV based on 2D CT slice images. It evaluates the performance of three peer-reviewed architectures (ResNet18, EfficientNetB0, ConvNeXt-Tiny), which have been fine-tuned to single-channel medical imaging. The training data included central CT slices from patient volumes in the TCGA-BLCA repository. The models benefited from stratified training and validation with class-weighted losses and data augmentation to ensure fair comparisons and clinical stability. The ConvNeXt-Tiny model achieved the highest validation accuracy of 98.5%, in addition to better pure precision and recall, among the 3 architectures. To improve the interpretability of the model and delineate stage-specific areas for the CT slices, Grad-CAM was utilized. The overall findings for SliceNet indicate the potential of light CNNs in support of accurate, real-time, and interpretable bladder cancer staging in the clinical workflow.