Prediction of Molecular Characteristics of Glioma by Machine Learning Using Only Hematoxylin-Eosin-Stained Images

Since the introduction of the concept of molecular classification, integrated diagnostics have become an essential tool, but this has led to the problem of high cost, and the requirement of a large amount of effort to reach a diagnosis. In this study, we investigated the possibility of performing integrated diagnoses of tumors using artificial intelligence (AI) to analyze only hematoxylin-eosin (HE)-stained slides. About 80 features of cell nuclei were analyzed, and using the support vector machine model and Random Forest algorithm to determine the expression of various molecules (isocitrate dehydrogenase 1, alpha thalassemia/mental retardation syndrome X-linked, p53 , and 1p19q co-deletion), tumor type, and WHO grading. Using this method, we were able to classify all tumor samples successfully. The present study suggests that the retrieval of tumor cell nuclear features from HE-stained slides by AI may contribute to the integrated diagnosis of glioma. However, 1p19q co-deletion was difficult to differentiate. Tumors with the 1p19q co-deletion were considered to have fewer nuclear shape changes than those with other gene mutations. Among such results, we found the possibility that this method could be used to search for the location at which the genetic mutation is acting on in the cell using the likelihood data.