Title:	Cell segmentation using machine and deep learning
DNr:	NAISS 2023/22-395
Project Type:	NAISS Small Compute
Principal Investigator:	Magnus Andersson <magnus.andersson@umu.se>
Affiliation:	Umeå universitet
Duration:	2023-03-30 – 2024-04-01
Classification:	10399
Keywords:

Abstract

Bacterial spores, also known as endospores, are dormant forms of sporulating bacteria that exhibit no cellular activity. Spores are exceptionally resilient to external stressors such as temperature, humidity, radiation, and chemical exposure. Due to their inherent resilience and ability to germinate back into bacteria when returned to more favorable conditions, spores from pathogenic bacteria pose a significant problem in many areas of society, including healthcare, food production, and homeland security. Therefore, studying spores is important for developing new sterilization and detection strategies. To study spores and determine morphology, size, ultrastructure, topography, and structural features, Transmission Electron Microscopy (TEM) can provide valuable information. In particular, TEM enables high-resolution visualization of all the layers within a spore, which for example can provide important clues of the mode of action of light or disinfection chemicals. However, the accurate segmentation and classification of bacterial spores are challenging and time-consuming tasks for researchers and microbiologists. The purpose of this proposal is to explore the application of machine learning and deep learning techniques to improve bacterial spore segmentation and classification accuracy. We want to develop an automated algorithm for segmenting and classifying layers in spore TEM images. The proposed algorithm will combine a CNN for feature extraction and an RF classifier for pixel classification. We will train the CNN with image data and use its predictions as input features to decision trees in the RF algorithm. Thus, the CNN will convert high-dimension 2D TEM images into low-dimension features that preserve the locality of pixels and reduce the curse of dimensionality for accurate prediction through the RF classifier.