Intelligent dual-modality label-free cell classification with light scattering imaging and Raman spectra measurements

Abstract

Leukemia diagnosis usually relies on multiple tests to analyze the morphological, biochemical, and genetic characteristics of blood cells. Accurate and label-free tools with the capabilities of interpreted cancer prediction and assorted diagnosing markers assure effective treatment; however, such an integrated system is currently lacking. In this study, we develop an intelligent dual-modality cell classification system with explainable machine learning by label-free measurements of the morphology features from elastic scattering images and molecular features from Raman spectra of single cells. For label-free identification of the normal granulocytes and two types of myeloid leukemia cells, the morphological features provide a classification accuracy of 96.04 % compared with 99.52 % using Raman molecular features. A high accuracy of 99.93 % is obtained by combining the morphological and molecular features. Here we reveal a total contribution of 82.49 %, 76.37 %, and 76.19 % for the morphological features compared with 17.51 %, 23.63 %, and 23.81 % for the molecular features for identifying the normal, CML, and AML cells, respectively. Integrating explainable machine learning with label-free dual-modality systems represents a promising tool for single-cell classification, which is foreseen to have great clinical applications by offering morphological and molecular information, reliable decision-making, accuracy, and automation.