Abstract
In diagnostic microbiology laboratories, manual cell counting using an optical microscope is an important test performed on several body fluid specimens. Manual cell counting is a laborious and slow process that requires a skilled microscopists to perform. This poses a challenge for microbiology laboratories as training and maintaining competency of microscopists at manual cell counting can be difficult, particularly for overnight staff or staff in peripheral laboratories that see less volume of specimens per day. Automated methods of cell counting such as flow cytometry have replaced the need for manual cell counting in some circumstances, however the manual method is still required in many contexts.
In the field of artificial intelligence (A.I), deep learning neural networks (DNNs) have become a powerful technology used for pattern recognition in a number of different ways. Convolutional neural networks (CNNs) were designed for object recognition in digital images, they have shown great value in many use cases for medical image analysis including in radiology, neurology, and cytology. Many large companies have now developed easy to use application programming interfaces (APIs) that allow a user to programme their own neural network models without requiring a background in computer science or programming; one such API is Custom Vision from Microsoft.
Techion are an information and technology company based in Mosgiel New Zealand, they specialize in microscopic imaging hardware, A.I machine learning software, and data management systems that allow diagnosis for parasite loads in faeces samples from stock animals. This integrated platform provides accurate and precise point of care results to customers without the need of a technician. Techion have recently developed a prototype microscope, the Nano-I, which has a more powerful microscope than its predecessors, which allows it to resolve smaller particles.
The aim of this project is to develop the Nano-I to take high resolution digital image captures of urine specimens collected from human patients. Once a large enough dataset of digital images is gathered that dataset was fed into a DNN which will be used to train an A.I algorithm to differentiate between different cell types in urine specimens.
Adjustments were made to how images were captured using the Nano-I as images were captured on samples mounted in a Neubauer counting chamber as opposed to the well system used in a Micro-I. Improvements were also made to the lighting of the microscope in order to achieve higher quality image captures. In total a dataset of 117 images were gathered using the Nano-I microscope and fed into Custom Vision to train a model to recognise three object categories: white blood cells (WBCs), red blood cells (RBCs), and epithelial cells. The overall object detector performance for the model across all three categories had precision of 90.1%, a recall 49.2%, and a mean average precision (mAP) of 75.1%. Of the three object categories, the model performed best at detecting epithelial cells with an average precision (AP) of 90.7%, then WBCs with an AP of 83.8% then RBCs with an AP of 50.8%.