Quantitative Nuclear Histomorphometry for distinguishing Undifferentiated Carcinoma from Non-Hodgkin’s Malignant Lymphoma and Benign Mimics

This is the authors' abstract. We don't add key points for this paper.

Background: Undifferentiated carcinomas visually cannot be discriminated from Non-Hodgkin’s Malignant Lymphomas (NHMLs) and need additional immunohistochemistry (IHC) examination. Some benign lesions are also mimicking the case above. Not every anatomical pathology laboratory is capable of performing IHC exams. Quantitative Nuclear Histomorphometry (QNH) is a computer-aided image analysis of digitized pathology image analysis. The aim of this study is to measure the accuracy of QNH to distinguish undifferentiated neoplasms and their benign as a potential for triagning IHC exams. Methods: Histopathological slides of undifferentiated carcinomas, NHMLs, and benign lesions that IHC has confirmed are collected. A series of images from selected regions of interest from Hematoxylin and Eosin (HE) slides are acquired under 100X and 400X magnifications. Hematoxylin-only images are generated from HE photomicrograph, further segmented, and its nuclear features are extracted. Features are further detailed in its five number summaries. One-vs-rest (OvR) lasso regression was performed to select relevant features. Other OvR binary logistical regressions are performed to create models for predicting diagnosis. The sensitivity, specificity, and accuracy of each prediction model are measured. Cohen’s kappa score measures interobserver reliabilities between model. Results: A sum of 31 cases were collected. A total of 344 and 670 PNG formatted of 1376 x 1038 pixels microphotographs from 100x and 400x magnification are acquired. Up to 300 thousand objects under 400x magnification and >1 million objects under 100x magnification are extracted. Benign lesions significantly (p<0.01) higher number of objects per image compared to both carcinomas and NHMLs. A total of 20 features for 100X magnification and 23 features for 400X magnification were selected based on Lasso Regressions. At image level, the accuracy of the model prediction under 100X and 400X magnification, each are 77.6% (p<0.01) and 83.6% (p<0.01). At slide levels in all magnification, model prediction accuracy of each magnification are (100X: 71.09%, 400X: 87.1%) compared to HE diagnosis accuracy (83.9%). Conclusion: The QNH can be used to predict and triaging undifferentiated carcinomas, NHMLs, and benign lesions before further examination. The QNH assessment also showed high reproducibility and more objective assessment in diagnosis.