As a major worldwide health issue, diagnosing colorectal cancer quickly and accurately is essential. The potential of deep learning to aid in colorectal cancer diagnosis has grown substantially in recent years. However, traditional deep learning models are frequently not widely adopted in clinical settings because to their high computational cost and resource intensiveness. To this end, we present "Deep Learning for Economical Colorectal Cancer Diagnosis: Harnessing the Power of Composite Networks with Unsupervised Learning," a novel method for analysing data in this area. The foundation of this study is the idea of using composite networks, which cleverly incorporates unsupervised learning to greatly lessen the training cost. Our algorithm effectively extracts relevant characteristics by exploiting vast volumes of unlabeled data and then fine-tuning on a smaller labelled dataset for targeted diagnostic tasks. The result is a cheap model that doesn't skimp on diagnostic precision. The empirical data we present here show that our composite network achieves high levels of accuracy while using a minimal amount of resources, making it a promising candidate for fast and cheap diagnosis of colorectal cancer. The results presented here highlight the potential of unsupervised learning to make deep learning more accessible and inexpensive for healthcare institutions around the world, and hence have far-reaching implications for the future of medical diagnosis.