Machine-learning-based classification of co-cultured cells from the phase images in digital holographic microscopy

Abstract

In the tumor microenvironment, cell interactions play a crucial role in influencing t he m orphology a nd metastasis characteristics of non-cancerous and cancerous cells. Although machine learning techniques have been proven effective i n c lassifying i ndividually c ultured c ell l ines, t heir a ccuracy m ay d eteriorate i n c lassifying a co-cultured mixture of cells. In the proposed work, the optical path difference images of human dermal fibroblast and melanoma A375 cells were recorded, individually and in a mixture, using the off-axis d igital holographic microscopy setup. A dataset consisted of segmented and labeled images of both sample types. The dataset was used to train a convolutional neural network through transfer learning to extract the morphology relevant features. An XGBoost classifier trained on the extracted features is found to effectively recognize morphological changes in cells within a mixture and classify them accurately with a limited size dataset.