Digitizing pathology lab workflows using image processing and OCR

Abstract

The pathology lab at Haukeland University Hospital is currently facing a few challenges. In the lab they process different specimens of tissue samples for various reasons such as looking for signs of cancers and tumours. The last six years the lab has had an average increase of 3.42% in the amount of test samples which needs to be analysed each year. This increase has led to queues forming at various stages of sample analysis. Specific samples are hard to locate within these queues and the queues lead to slowdowns of sample processing. The pathology lab require a better solution to the way and order in which samples are processed and tracked, to do so they need to gather data about the processing by implementing a tracking solution. This project aims to help them achieve this goal by looking for a potential software solution to part of the problem. This solution aims to take advantage of technologies such as optical character recognition (OCR) for detecting and tracking samples. The goal for this research is to create and test solutions for cassette detection and identification using pre-trained image processing libraries. Testing two different methods, these being edge detection and the EAST neural network, they achieved an accuracy of 77.84% and 93.41% respectively regarding cassette detection. Tesseract OCR performance of detected cassettes also varies between the two methods, giving an accuracy score of 36.1% when using edge detection and 62.1% using EAST. The increase in accuracy comes at a cost in runtime. In addition to these evaluations an in-lab trial compares the sorting time for the current solution of manual sorting versus the efficiency of sorting using the proposed digital solution. The trial concluded that the proposed digital solution is able to increase the amount of cassettes sorted within a set amount of time by 54% decreasing the time spent on manual sorting activities by 35%. This thesis also covers some of the interaction design decisions for the proposed application to allow for manual error correction. Through conceptual designs the thesis shows how the proposed system could interface with a process execution engine. There is also a proposal for what the architecture of an integrated system could look like. Integrating this system would allow for the generation of fine-grained event logs for process mining purposes. The data from these logs have a possibility of leading to future improvements in pathology lab workflow.