Deep learning-based object detection for soil bacterial community analysis in microfluidics

Microfluidics is a multidisciplinary platform that integrates microfabrication, physical chemistry analysis, automation, and microscopy. It has the advantages of precise liquid manipulation, rapid measurements, and real-time visualization at the microscale, which is especially of interest and benefit to microbial studies. Soil Chips are microfabricated microfluidic devices typically made of glass and polydimethylsiloxane (PDMS), designed to mimic the real soil network and allow real-time visualization and characterization of microbial activity at the micro-scale. They have so far been used to investigate microbial activities, interactions, community composition, and distribution under different conditions in soil analog systems. Challenge comes when working with natural soil samples. Due to mineral aggregates and debris, valuable information such as the abundance of individuals, cell morphology, and the relationship between bacteria and their geochemical and physical environment are difficult to extract via a simple thresholding method. Since microorganisms and microfluidic structures have distinct features from the noisy background that can be easily picked up by our eyes, a biologically inspired convolutional neural network model for object detection is the most suitable tool for this task.
We used a small part of data from three different experiments to train a well-developed object detection and segmentation algorithm Mask RCNN and implemented further analysis of bacteria abundance, spatial distribution, and morphological characterization. We are able to plot the distribution of all the detected bacteria including clusters in terms of abundance, size, shape, and index of aggregation. A distinct difference in bacteria characteristics can be observed in the samples acquired from three locations (Greenland, Sweden, and Kenya). We are now planning to extend the classification library to include other microbial groups including fungi, protists, invertebrates, and micro arthropods.