Embedded Files
Figure 1. An example of classifying blood cell types via manul gating [1].
Background
Background
Leukemia is a type of cancer that affects the blood and bone marrow. The major types are acute leukemia (i.e. ALL) and chronic leukemia (i.e. CLL), which are grouped by how fast they develop and which blood cells are evolved. One of the most commonly used diaognistic methods is flow cytometry immuno-phenotyping. It detects markers specific to blood cells or bone marrow cells in normal and abnormal individuals, allowing physicians to identify the specific type of leukemia and prepare for corresponding treatment.
However, the flow cytometry immuno-phenotyping in hospitals is done primarily through manual gating, which is time-consuming and prone to human errors.
Our Approach
Our Approach
Figure 2. Example templates of classifying blood cell types via our UMAP pipeline. The red points and blue points on the leftmost plot represents the cells from ALL patients and cells from non-ALL reference samples, respectively. The rightmost plot is the result of classification via HDBSCAN, an algorithms that can be applied to UMAP layouts.
Fortunately, the current progress in machine learning and deep learning algorithms shields light on the analysis of high-dimensional dataset like flow cytometry data, so we can look at multiple markers simultaneously and make a consistent and justifiable prediction for patients.
Instead of finding boardlines on a 2-Dimensional space, we utilizes machine learning algorithms, UMAP (Uniform Manifold Approximation and Projection for Dimension Reduction), as well as the CNN (Convolutional Neural Network) model to analyze the given multi-dimentional flow cytometry dataset.
For ALL patients, we
1) train multiple UMAP samples using known patients data, and
2) make predictions via machine learning algorithms for new data under single-positive voting procedure.
For CLL patients, we
1) generate several UMAP templates per sample first and then,
2) train them using CNN models to predict results.
Via further training and testing, we demonstrate that these approaches have the potential to make the diagnosis of Leukemia faster and more reliable.
Page Leader: Shihui Zhu
Page updated
Report abuse