Hematoxylin and eosin (H&E) staining is a commonly used technique for diagnosing cancer in histopathological slides. However, it has limitations such as time-consuming procedures and analyzing samples in two dimensions. To address these issues, we propose integrating refractive index (RI) tomography and deep learning. RI tomography is a label-free imaging technique that provides quantitative three-dimensional (3D) RI distribution without external staining. We trained a neural network to map RI images to colored bright-field (BF) images using paired samples of colon cancer slides. The network was then used to predict BF images from label-free RI images, generating 3D virtually stained H&E images. The accuracy was verified by comparing them to conventionally stained images. This approach combines deep learning with RI tomography to virtually stain subcellular information from RI distributions, enabling a deeper understanding of histopathological slides and improving cancer diagnosis.