[P-23] Accuracy of nuclear features granulometry to predict malignant lymphomas from undifferentiated carcinoma and lymphocytic benign lesions

Okky Husain^1,2, Hermin Aminah Usman^1,2, Hasrayati Agustina^1,2, Widyawardana Adiprawita³ and Muhammad Syah Misuari Sabirin^1,2

Background and Objectives: Haematoxylin and eosin (HE) morphology of lymphomas and undifferentiated carcinomas are challenging to be distinguished. Ancillary phenotype examinations by immunohistochemistry (IHC) are commonly used however not readily available. Granulometry is mathematical approach to extract objects size distribution from digital image. The aim was to predicts lesion based on HE nuclear granulometry.

Materials and Methods: Digital images of 100x and 400x magnifications from HE slides of IHC confirmed malignant lymphoma, not otherwise specified; undifferentiated carcinoma; and lymphatics benign lesions. Area (linear and geometrically), and intensity (maximum, minimum, range, and means) of nuclei from each Images were automatically measured using scikit-image measure module. Nuclei samples were randomly pooled unto 5 arrays of equal magnification and cases and compared to each images using unpaired parametric statistical t-test.

Results: Under 100x and 400x, there were total images of 121 and 208 from lymphomas, 150 and 254 from carcinomas, and 100 and 215 from benign lesions. Images dimensions were 1,376 x 1,038 pixels, 24-bits depth (8 bits, 3 channels). Under 100x and 400x, nuclei intensity accuracy for lymphomas were 50% and 76%, carcinomas were 57% and 85%, benign lesions were 52% and 65%. Under 100x and 400x, nuclei area accuracy for lymphomas were 50% and 65%, carcinomas were 57% and 70%, benign lesions were 53% and 61%.

Conclusion: Nuclear HE granulometry intensity range is the most helpful to distinguish lymphomas from carcinoma and benign lesions under 400x magnification.