Abstract
For patients with acute ischemic stroke, histological quantification of thrombus composition provides evidence for determining appropriate treatment. However, the traditional manual segmentation of stained thrombi is laborious and inconsistent. In this study, we propose a label-free method that combines optical diffraction tomography (ODT) and deep learning (DL) to automate the histological quantification process. The DL model classifies ODT image patches with 95% accuracy, and the collective prediction generates a whole-slide map of red blood cells and fibrin. The resulting whole-slide composition displays an average error of 1.1% and does not experience staining variability, facilitating faster analysis with reduced labor. The present approach will enable rapid and quantitative evaluation of blood clot composition, expediting the preclinical research and diagnosis of cardiovascular diseases.
Keywords : acute ischemic stroke; optical diffraction tomography; thrombus composition; label-free; deep learning
Abbreviations: ODT , optical diffraction tomography;DL, deep learning; RBC , red blood cells; BF , bright-field; RI , refractive index; DMD , digital micromirror device; FOV , field of view; CatSIM , categorical image similarity metric; ROSE , rapid on-site evaluation
1 INTRODUCTION
Acute ischemic stroke (AIS) is a fatal disease that requires immediate and appropriate treatment to prevent death or devastating sequelae. The composition of the thrombus, or blood clot, provides useful evidence in determining the appropriate treatment for AIS. The histological composition and structure of AIS thrombi offer insights into disease pathophysiology, such as pre-interventional migration, vascular origin, and thrombus age, and may be used to decide the best treatment options for patients. The thrombus composition is also related to the clinical outcomes of recanalization with thrombolysis or endovascular thrombectomy.[1] Studies have identified the pre-interventional migration, vascular origin, and age of thrombi from their histological composition and structure,[2,3]while others have differentiated cardioembolic from non-cardioembolic origin using the proportion of fibrin/platelets or red blood cells (RBCs) in thrombi.[4,5]
Conventional approaches to evaluating the histological composition of a thrombus require pathologists to manually screen stained thrombus sections under a bright-field (BF) microscope. Through microscopic examination, pathologists differentiate the type of thrombi into RBC-dominant or fibrin/platelet-dominant and report distinct features based on the stained thrombus slide.[6-8] However, color-based analysis of a thrombus slide is highly influenced by the staining quality and may lead to generalization problems due to staining variability.[9] Additionally, the fixation and staining procedures are labor-intensive and time-consuming, limiting efficiency and throughput.
In this study, we propose and experimentally demonstrate a label-free histological quantification framework to rapidly assess AIS thrombi using optical diffraction tomography (ODT) and a deep learning (DL) algorithm (Figure 1 ). We leverage the label-free structural assessment of ODT with the statistical image recognition of deep learning to automatically characterize a whole unstained slide of the thrombus. ODT is a label-free 3D imaging method for cells and tissues that returns the refractive index (RI) distribution of the sample.[10,11] The label-free nature of ODT simplifies sample preparation and enables efficient imaging, while also providing the ability to extract various quantitative biophysical properties from the RI distribution.