4 DISCUSSION
This study presents a DL-based approach to classify the histological composition of thrombi using label-free ODT images. The DL model achieved over 95% accuracy in classifying patches into RBC, fibrin, and background, enabling the prediction of thrombus composition at whole-slide level. The use of ODT provided higher resolution and contrast compared to conventional BF microscopy followed by H&E staining. This label-free approach also allowed for rapid histological analysis without staining and reduced the variability of staining-dependent color distribution. The present work demonstrates the rapid label-free detection a spatial distribution map of thrombus composition that can be used to assess the thrombus response to thrombectomy procedure.[39]
The results shown here can also be utilized for in-depth investigation of a thrombus slide. The RI contrast and subcellular spatial resolution provided by ODT facilitate a more detailed assessment of thrombus structures compared to BF microscopy followed by H&E staining. Pores within the fibrin network were more clearly observed in ODT than in BF images. The border and texture of individual RBCs also appeared sharper in ODT. Additionally, ODT distinguished structures that were not differentiated by BF microscopy, such as regions of compressed RBCs tentatively identified as polyhedrocytes. Polyhedrocytes may lead to more thrombolysis-resistant thrombi by forming an impermeable layer that blocks fibrinolytic enzymes from diffusing inside the thrombi. These observations indicate that ODT measurement could potentially identify indicators of thromboembolism.[40]
Additionally, accounting for volumetric information could provide a more accurate determination of thrombus composition due to intra-thrombus heterogeneity. Ground truth generation could also be improved by utilizing immunofluorescence or immunohistochemistry. The DL design could be modified to carry out segmentation instead of patch classification for better reflecting the high-resolution geometry provided by ODT. Using identical slides for ODT and bright-field imaging may also improve pixel-level registration and provide more accurate supervised learning for segmentation networks.
The proposed framework has the potential to complement the current clinical routine by providing rapid on-site evaluation (ROSE) of thrombi. H&E staining and inspection of thrombi remains a solid gold standard in most clinical institutes owing to the effectiveness that results from abundant domain knowledge. Our noninvasive test can be integrated into this routine slide inspection workflow without perturbing or prohibiting the existing processes. The use of label-free ODT enabled us to reduce the time and associated effort of staining. Our approach does not require any staining and can directly assess the unlabeled specimen, allowing rapid histological analysis of patients with AIS. Our approach can also provide consistent results compared with conventional methods that suffer from staining variability and interpreter fatigue.
The conventional composition assessment by practitioners depends heavily on the color distribution which results from staining.[7] As the staining procedures rely on human or environmental conditions, the color distribution may be significantly different.[9,41,42] By contrast, it does not apply to our label-free approach that measures RI, an intrinsic physical quantity of the native sample.
There are technical limitations in this work, such as the bulkiness and complexity of the ODT hardware implementation, the limit in imaging speed arising from the small FOV compared to a whole slide, and the acquisition rate of the image sensor. However, recent developments in non-interferometric ODT using low-coherence sources and engineering approaches such as multiplexing can mitigate these issues.[43,44] Other possible engineering approaches include multiplexing which expands the measurement bandwidth by exploiting the polarization[45] or spectral dimension.[46]
5 CONCLUSIONS
We proposed and experimentally demonstrated a rapid and fully automated prediction of thrombus composition using label-free refractive index (RI) tomography and deep learning. Our proposed convolutional neural network model accurately classified thrombus subtypes for each label-free RI patch with high accuracy (>95%), affording quantitative analysis of unstained thrombus slides. We believe that this approach, which does not require any staining or manual inspection, will significantly accelerate thrombus biopsy procedures for diagnosing ischemic stroke and related diseases.
In future work, we may extend our framework to fully exploit the 3D information of the RI tomogram, which could enhance the accuracy of the thrombus composition inference. Our DL-based approach using label-free ODT images presents a promising alternative to conventional staining-based methods for histological analysis of thrombi, which can also be further extended to identify and classify cell types in tissue slides. The proposed framework has the potential to provide more efficient and consistent analysis while complementing the existing clinical routine.