Phenotyping neuroblastoma cells through intelligent scrutiny of stain-free biomarkers in holographic flow cytometry

Abstract

To efficiently tackle certain tumor types, finding new biomarkers for rapid and complete phenotyping of cancer cells is highly demanded. This is especially the case for the most common pediatric solid tumor of the sympathetic nervous system, namely, neuroblastoma (NB). Liquid biopsy is in principle a very promising tool for this purpose, but usually enrichment and isolation of circulating tumor cells in such patients remain difficult due to the unavailability of universal NB cell-specific surface markers. Here, we show that rapid screening and phenotyping of NB cells through stain-free biomarkers supported by artificial intelligence is a viable route for liquid biopsy. We demonstrate the concept through a flow cytometry based on label-free holographic quantitative phase-contrast microscopy empowered by machine learning. In detail, we exploit a hierarchical decision scheme where at first level NB cells are classified from monocytes with 97.9% accuracy. Then we demonstrate that different phenotypes are discriminated within NB class. Indeed, for each cell classified as NB its belonging to one of four NB sub-populations (i.e., CHP212, SKNBE2, SHSY5Y, and SKNSH) is evaluated thus achieving accuracy in the range 73.6%-89.1%. The achieved results solve the realistic problem related to the identification circulating tumor cell, i.e., the possibility to recognize and detect tumor cells morphologically similar to blood cells, which is the core issue in liquid biopsy based on stain-free microscopy. The presented approach operates at lab-on-chip scale and emulates real-world scenarios, thus representing a future route for liquid biopsy by exploiting intelligent biomedical imaging.

FIG. 1.

Holographic imaging flow cytometry for distinguishing several subtypes of NB cells (i.e., CHP212, SKNBE2, SHSY5Y, and SKNSH) from monocytes. (a) Sketch of the holographic imaging flow cytometer. HWP: half-wave plate; PBS: polarizing beam splitter; L1, L2: Lens; M: mirror; MO: microscope objective; MC: microfluidic channel; TL: tube lens; BS: beam splitter; CMOS: camera. (b) Digital hologram recorded by the holographic imaging flow cytometer, with cells imaged while flowing along the y-axis and rotating outside the image plane. (c) Overall pipeline of the proposed strategy. $N$ QPMs ( $200\times 200$ square pixels size, 5 μm scale bar) for each cell are numerically retrieved from the recorded digital holograms, and 37 features are measure for each of them. The extracted features are fed to a hierarchical classifier (see the sketch with fake class names). The $N$ predicted outputs are used to infer the cell line of the analyzed cell by means of max-voting.

FIG. 2.

Inspection of the dataset collected by the holographic imaging flow cytometer. (a)–(c) Box plots about some conventional features, i.e., dry mass, area, and GLCM energy, respectively, computed from the QPMs for each single cell. (d)–(f) Box plots about some fractal features, i.e., fractal dimension, lacunarity index, and regularity index, respectively, computed from the QPMs for each single cell. (g) Scatter plot of the first two PCA components computed from all the conventional and fractal features about monocytes and NB cells. (h) Scatter plot of the first two PCA components computed from all the conventional and fractal features about the four NB subtypes. (i) Sketch of the hierarchical classifier made of three levels (L1, L2, and L3) and four single classifiers (L1, L2, L3.1, and L3.2). The intermediate NB1 class includes CHP212 and SKNBE2 cells [gray circle in (h)]. The intermediate NB2 class includes SHSY5Y and SKNSH cells [cyan circle in (h)].

FIG. 3.

Representation of the training sets by means of the t-SNE algorithm for each classification problem (rows) and for each feature set (columns). (a)–(c) Training set for discriminating monocyte vs NB cells by means of 24 conventional features, 13 fractal features, and 37 hybrid features, respectively. (d)–(f) Training set for discriminating NB1 vs NB2 by means of 24 conventional features, 13 fractal features, and 37 hybrid features, respectively. (g)–(i) Training set for discriminating CHP212 vs SKNBE2 by means of 24 conventional features, 13 fractal features, and 37 hybrid features, respectively. (j)–(l) Training set for discriminating SHSY5Y vs SKNSH by means of 24 conventional features, 13 fractal features, and 37 hybrid features, respectively.

FIG. 4.

Classification performances within the hierarchical model. (a) and (b) Recall and accuracy, respectively, computed over the QPMs of the test set without max-voting by using the conventional features, fractal features, and hybrid features. (c) and (d) Recall and accuracy, respectively, computed over the cells of the test set through max-voting by using the conventional features, fractal features, and hybrid features. (e) Sketch of the max-voting strategy. For each cell flowing along the y-axis and rotating outside the image plane, N QPMs are recorded. For each QPM, its features are extracted to feed a shallow neural network and predict its class (A or B). The cell is assigned to the class that has occurred more times (N_A > N_B or N_B > N_A). (f) Sketch of the hierarchical model along with the best performances of each classifier obtained after using the reported feature sets combined to max-voting.

FIG. 5.

Classification and QPM visualization of several cell lines by means of holographic flow cytometry. (a)–(e) Global probabilities P of correctly classifying monocytes and CHP212, SKNBE2, SHSY5Y, and SKNSH NB cells, respectively, obtained after multiplying the recall values found along each corresponding path inside the best hierarchical tree. (f)–(j) QPMs taken from dataset exploited for the max-voting-based classification of the cell lines in (a)–(e), respectively. Scale bar is 5 μm.

FIG. 6.

2D images used for feature extraction. (a) QPM ( $200\times 200$ square pixels) of an SHSY5Y cell (yellow) and its zero-padded $256\times 256$ version. (b) Support map obtained by segmenting the zero-padded QPM in (a). (c) Gradient magnitude of the zero-padded QPM in (a) normalized to its maximum value. (d) Hole support map obtained by applying a 0.3 thresholding to the normalized gradient magnitude in (c). Scale bar is 5 μm.