Development of AI assisted light microscopy system for blood cell analysis and infection detection

Abstract

Differential counting for quantification of blood cell sub-populations and assessment of blood cell status is crucial for diagnosing and monitoring various diseases and conditions. Blood comprises three main components: plasma, erythrocytes, and leukocytes. Morphological analysis of these components, through techniques such as blood smears and flow cytometry, enables the identification of various haematological conditions and infections.

This research aimed to develop an artificial intelligence (AI) assisted system that integrates light microscopy, microfluidics, and deep learning models to automate and enhance blood cell quantification and leukocyte recognition. By addressing the limitations of traditional methods, such as manual evaluation and inter-observer variability, this system seeks to provide accurate, efficient, and standardised analysis of blood samples. The integration of AI and light microscopy promises to revolutionise haematological diagnostics, enabling more precise and accessible multi-pathology diagnostics.

First, a microfluidic microscopy platform capable of high-throughput quantification and morphological analysis of blood cell populations has been developed and tested to validate iterative improvements. This prototype combined automated sample handling, data recording, and advanced computational techniques, offering a robust solution for comprehensive blood cell analysis. Then, the built device was used to prove the robustness of image acquisition to train a You Only Look Once (YOLO) artificial neural network for blood cell detection and classification. The generated models used the images acquired through the custom-designed device to identify erythrocytes, thrombocytes and leukocyte subtypes automatically. The AI models underwent validation using curated test datasets and comparison to gold standard research, achieving an accuracy of 87.6% compared to ground truth. Comparison tests with cytometry techniques using flow cytometry and haematology analysers confirmed the correct functioning of the developed AI-based blood cell subtype identification tool. Finally, building upon the previous work, the neural network was expanded to identify activated leukocytes. The trained neural network was able to discriminate between activated and non-activated lymphocytes and neutrophils with 82.0% accuracy for the full leukocyte panel and 90.0% accuracy for the monocyte-only network. Proving its ability to distinguish activation of leukocytes when in the presence of infection.

The presented research proves the ability of the final prototype device to perform a full blood count, with comparative accuracy to current healthcare haematology analysers, while performing a blood smear-like analysis in one step. The research also demonstrates the ability to differentiate leukocytes when activated. This research sets the scene for the creation of a sample multiple assessments devices combining microfluidic imaging with AI-driven diagnostic analysis by including other follow-up analyses after the blood differential count, incorporating the detection of leukaemia or anaemias, for example.