An Intelligent Image-based Colourimetric Test Framework for Diagnosis

The global imbalance between the healthcare provider and patient ratio, an increasingly elderly population and resource-limited settings have triggered the demand for point-of-care (POC) platforms, prompting the growth of personalised healthcare and homecare solutions. This thesis presents an investigation into an AI-enabled image-based system to perform automatic colourimetric tests in real-time. The case study of wet-chemical-based enzyme-linked immunosorbent assay (ELISA) and dry-chemical-based lateral flow assay (LFA) were utilised to design and develop an intelligent framework for chromaticity analysis with minimal user intervention or additional hardware attachments. The proposed system was designed by exploring state-of-the-art solutions for each component of an image-based colourimetric test, trial and error, and domain knowledge. At first, a reaction phase and time-dependent approach was proposed to track the dynamic changes in a colourimetric reaction by calculating the Euclidean distances. Subsequently, the final static stage of the reactions were considered and the images were pre-processed and segmented before applying vigorous noise removal techniques. The 10-fold cross-validated classifiers were trained with the optimum number of features using supervised machine learning. A completely separate testing dataset was utilised while testing the model. Additionally, a pre-trained model of deep learning was deployed to determine the type of colourimetric test, which can be integrated into the system where feasible. Based on our study, the reaction phase and time-dependent scheme was found to be more suitable for wet-chemical-based assays, particularly for low concentration samples. In addition to classification, the approach can assist in optimising the reaction time. However, due to the requirement of significant memory space by the video frames, the final system consisted of an alternative approach - considering only the reaction phase and time-independent scheme. On an ideal condition, the later approach provided more than 98% accuracy for colourimetric decision. Furthermore, the exploration of a pre-trained deep learning model revealed its strength in the test-type detection, instead providing the colourimetric classification. Therefore, deep learning was deployed to initiate the system based on the assay type (i.e. ELISA or LFA), which provided 100% accuracy. The system we demonstrated complies with the ASSURED criteria. As compared to the existing systems, the proposed intelligent and robust system with real-time processing capabilities has experienced a more extensive course of validation to enumerate the reliability of the system. Unlike most of the works in the literature, the proposed system provided the colourimetric prediction without any opto-mechanical attachment. Such an easy-to-use and computationally efficient system can be integrated into a server or deployed on a mobile platform to create better harmony between biochemical and computational complexity and eliminate the subjectivity of interpretation.