Text

theses

ETD graduate

Advancements in computer vision methodologies and machine learning in the medical domain have played a major role in diagnostics and clinical pathology. Cell quantification from whole blood can aid in detecting and managing infections, cardiovascular diseases and biomarker detection which in turn helps in understanding the immunological and genetic disorders, cancers, etc. Developing a point-of-care solution for this will accelerate the therapy timeline and increase the accessibility across the world. Our lab has previously developed a smartphone based microfluidic biosensor for capturing the microscopic images of various components of the blood cells. Using this design, in this study, a deep learning-based cell quantification from the captured images is investigated and the cell counts are predicted using a convolutional neural network architecture. The proposed methodology was evaluated on a dataset varying in numbers, clarity, smartphones, fluorophores and cell numbers. This model was then integrated into an Application Programming Interface (API) to predict the cell counts from an image using the trained model. Our results showed successful prediction of cell counts from a smartphone captured image in cross-validation with R2 = 0.99 for N=33. This helps in eliminating the need for manual pre-processing of an image and morphological methods for cell counting which is a user-skill based approach. This proposed Deep Learning based cell quantification has shown agility and more automated process when compared to the benchmark techniques.

ETD_11448

M.S.

Includes bibliographical references

doi:10.7282/t3-0mc4-jd08

The author owns the copyright to this work.