Early cancer detection continues to be a significant challenge in clinical practice due to limitation of conventional diagnostic technique that often takes time and error prone. This systematic review evaluates the efficacy of deep learning (DL) architecture and datasets to improve cancer detection and diagnosis. We performed a structural analysis on 40 high-impact research paper published in Q1 journals between 2014 and 2025, considering DL model performance, datasets, and clinical relevance. Results indicate that fundamental architectures such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) consistently report high diagnostic accuracy (>90%) on radiology- and histopathology-based imaging datasets. Conversely, DL performance on non-imaging clinical data, including electronic medical records (EMDs), is more varied. Evaluation metrics such as AUC and DICE shows the trade-off between classification precision and segmentation accuracy. Despite their potential, DL models have significant limitations in terms of generalization, interpretability, and integration within real-world clinical workflows. This review highlights the need for standardized evaluation, implementation of ethical models, and multi-modal data fusion to facilitate wider and more equitable clinical uptake of DL in cancer diagnostics.