Clustering and Deep Learning have garnered significant attention in recent years, particularly through unsupervised neural networks like autoencoders, which automatically uncover data structures. A key link exists between clustering and representation learning: effective representations lead to better clustering, and conversely, clustering aids in representation learning. However, many approaches rely on the K-means framework, which requires predefining the number of clusters (K). To address this, the DipDECK method uses the Dip-test to determine K. This paper introduces a modified version, Mod-DipDECK, which enhances the embedding space, allowing clearer cluster structures by applying an orthonormal transformation matrix derived from the intraclass scatter matrix of K-means. Our research eliminates the decoder and employs a greedy optimization approach, demonstrating that Mod-DipDECK outperforms state-of-the-art algorithms in real-world datasets.