Files
Abstract
Standard clotting tests usually do not suffice in investigating the mechanisms of thrombosis in COVID-19 patients. Flow cytometry provides a superior alternative since it enables research on sub-cluster cellular components from blood samples. Manual gating has conventionally been used to detect and analyze such sub-clusters, but it is labor-intensive and prone to subjective mistakes on behalf of human experts. This work proposes a deep learning-based method that automatically identifies and classifies cellular clusters to facilitate more efficient analysis of immuno-thrombosis. The method consists of two stages. The first stage employs a customized convolutional neural network (CNN) to classify grayscale images into clusters and non-clusters, with a test accuracy of 87.5%. The second phase focuses on multi-cell cluster images, using pre-defined color-based criteria to identify sub-cluster elements and obtain 85.6% accuracy. In addition, the viability of few-shot learning using ChatGPT 4o is explored, followed by a comparison with the customized CNN, the pre-trained deep learning models, and machine learning classifiers. The customized CNN outperformed the ML classifier and ChatGPT 4o and was outperformed by pre-trained deep learning models. Grad-CAM visualizations were used to perform a misclassification analysis of the CNN-based model predictions to enhance the interpretability of test results.