Fumikazu Kimura

Immunohistochemistry and machine learning study of DNA replication-associated proteins in uterine epithelial tumors and precursor lesions

Endometrioid adenocarcinoma (EA) has been on the increase in recent years in developed countries. Early detection of endometrioid adenocarcinoma in the endometrial corpus is crucial for patient prognosis and early treatment, although their distinction can sometimes be challenging. In this study, we focused on DNA replication-related proteins through immunohistochemical analysis and investigated whether the discrimination between EA and their precursor lesions is achievable using machine learning techniques. The research utilized tissue specimens from 100 cases, including EA of different grades (Grade 1; G1, Grade 2; G2, Grade 3; G3) and their precursor lesions (endometrial hyperplasia without atypia; EH, endometrial atypical hyperplasia: AH). Immunohistochemical analysis of DNA replication-related proteins, such as ORC1, Cdt1, Cdc6, MCM7, Cdc7, and Geminin, was conducted for each case, measuring the Labeling Index (LI) and optical density (OD) of protein expression. Furthermore, we performed statistical significance tests and machine learning -discriminant analysis using LI and OD as inputs, employing non-linear Support Vector Machines (NSVM). The NSVM discriminant analysis demonstrated the accuracy of over 85 % between EH and each differentiation grade of EA, the accuracy is also similar for AH and each differentiation grade of EA. In addition, changing the combination of DNA replication-related proteins used for discrimination resulted in a high accuracy (95-100 %). A discriminant analysis with NSVM using the LI and OD of DNA replication-related proteins may enable the differentiation of EA from its precursor lesions.

Discriminant analysis and interpretation of nuclear chromatin distribution and coarseness using gray‐level co‐occurrence matrix features for lobular endocervical glandular hyperplasia

AbstractBackgroundLobular endocervical glandular hyperplasia (LEGH) is a disease considered to be the origin of tumorigenesis of minimal deviation adenocarcinoma, which has characteristic expression in the gastric pyloric mucosa. It is difficult to diagnose by nuclear findings because of lower nuclear atypia. In this study, nuclei of endocervical (EC) and LEGH cells were digitized, and nuclear information was quantified from nuclear images and objectively evaluated using a computer. We examined whether it is possible to distinguish between EC and LEGH cells, which is difficult by human eyes.MethodsSignal intensity, morphological features, Otsu thresholding technique and gray‐level co‐occurrence matrix (GLCM) features were calculated from nuclei of EC and LEGH cells on cytology microscopic images. Then, discriminant analysis was performed using the significant difference test and linear support vector machine (LSVM).ResultsGLCM features in LEGH cells were higher than those in EC cells. The nuclei of LEGH cells had a higher frequency of signal value pairs with a larger signal value difference than that of EC cells. Therefore, LEGH cell nuclei are thought to have more chromatin granules, and the chromatin is coarse and granular. Moreover, in the LSVM discriminant analysis, the accuracy of GLCM calculated using these features was 85.4%.ConclusionIn this study, GLCM accurately demonstrated the nuclear chromatin distribution and coarseness. Discriminant analysis of EC and LEGH cells using GLCM features is useful.