Point-of-care diagnostics with digital microscopy and artificial intelligence

The lack of access to diagnostics is a global problem which causes underdiagnosis of various common and treatable diseases. In certain areas, the access to laboratory services and medical experts is extremely limited, such as in sub-Saharan Africa, with often less than one practising pathologist per one million inhabitants. Annually, hundreds of millions of microscopy samples are analysed to diagnose e.g. infectious diseases and cancers, but the need for more is significant. During the last decade, technological advancements and reduced prices of optical components have enabled the construction of inexpensive, portable devices for digitization of microscopy samples; a procedure traditionally limited to well-equipped laboratories with expensive high-end equipment. By allowing digitization of samples directly at the point of care (POC), advanced digital diagnostic techniques, such as the analysis of samples with medical ‘artificial intelligence’ (AI) algorithms, can be utilized also outside high-end laboratories – which is precisely where the need for improved diagnostics is often most significant. The aim of this thesis is to study how low-cost, POC digital microscopy, supported by automatized digital image analysis and AI can be applied for routine microscopy diagnostics with an emphasis on potential areas of application in low-resource settings. We describe, implement and evaluate various techniques for POC digitization and analysis of samples using both visual methods and digital algorithms. Specifically, we evaluate the technologies for the analysis of breast cancer tissue samples (assessment of hormone receptor expression), intraoperative samples from cancer surgeries (detection of metastases in lymph node frozen sections), cytological samples (digital Pap smear screening) and parasitological samples (diagnostics of neglected tropical diseases). Our results show how the digitization of a variety of routine microscopy samples is feasible using systems suitable POC usage with sufficient image quality for diagnostic applications. Furthermore, the findings demonstrate how digital methods, based on computer vision and AI, can be utilized to facilitate the sample analysis process to e.g. quantify tissue stains and detect atypical cells and infectious pathogens in the samples with levels of accuracy comparable to conventional methods. In conclusion, our findings show how technological advancements can be leveraged to create general-purpose digital microscopy diagnostic platforms, which are implementable and feasible to use for diagnostic purposes at the POC. This allows the utilization of modern digital algorithms and AI to aid in analysis of samples and facilitate the diagnostic process by automatically extracting information from the digital samples. These findings are important steps in the effort to develop novel diagnostic technologies which are usable also in areas without access to high-end laboratories, and the technologies described here are also likely to be applicable for diagnostics of other diseases which are currently diagnosed with light microscopy.