Over the last decades, a host of gene expression profiles of tumor and normal tissue samples have been recorded by many microarray and RNA-Seq projects. Much of this big data awaits a full understanding and exploitation for translational cancer research. In particular, the pressing need to discover gene panels for diagnosis and therapy have not received yet a definitive answer. Here, we tackle such a question through rigorous mining of some of the currently available data. Our mining scheme rests on formal concept analysis and rough set theory and allows us to identify perfect gene panels for twelve of the solid tumors reported in the TCGA database. We dub them 'perfect gene panels' because they perfectly discriminate between normal and tumor samples. To wit, testing the gene expression profiles against a tumor or normal pattern provides no false positive and no false negative cases (i.e., 100% sensitivity and 100% specificity). Hence, perfect gene panels might be useful genetic markers for cancer diagnosis. Furthermore, we stress that such panels come in many flavors depending on the gene expression levels we choose as a pattern to check. For instance, there are perfect panels where a single gene over-expression signals a tumor and others where a single non-silenced gene is an indication of a tumor-free sample, just to mention two out of eight possible cases. Remarkably, some panels also suggest suitable genetic targets for therapeutic interventions, since they define normal samples by tuning the expression level of a single gene.