Electron tomography (ET) is a powerful tool for the 3D mapping of the complex 3D sub-cellular structures of cells. It can provide detailed structural data for the extraction, segmentation and annotation (e.g. organelle type, spatial location, volume, surface area, shape and cellular interactions). The ability to map and model sub-cellular volumes is dependent on the quality of the electron tomography data and the ability to segment the resolved features accurately. In particular low signal to noise ratios can resolute in the loss of structural data as well as the incorrect identification of false positives. Manual segmentation is currently considered the gold standard but is subjective and the process is slow. This is highlighted by the finding that the careful segmentation of ~1% of an insulin-secreting HIT-T15 cell required approximately 3600 person-hours (Marsh et al., 2001a). Consequently as the volume and quality of cellular electron tomography data increases so will the need for automated segmentation approaches. Such automated processes will likely require the integration of image filtrations methods, boundary-based and region-based segmentation algorithms and edge detector algorithms. To be of real utility these automated methods must be fast as well as accurate ideally across multiple scales ranging from tissues to molecules. Semi-automated approaches will also be of value if they are able to yield significant gains in data quality and speed.
The process of segmentation is principally made difficult by limitations caused by low signal-to-noise ratio (SNR) of volumetric image data typical of that generated by electron tomography. Indeed compared with MRI and CT data-sets electron microscopy has a low SNR and so is good test system for the development of segmentation algorithm. To date the low SNR of electron tomography data has resulted in limited examples of successful automatic segmentation. Improved image pre-processing techniques, such as increasing the SNR through improved sample preparation and imaging, as well as the careful application of denoising algorithms in conjunction with carefully managed segmentation processes have proven most beneficial, but there is still substantial scope for improvement.
The aim of this project has been to analyse the pancreatic beta cell tomograms and to conduct a detailed investigation into the structural diversity of their insulin granules, mitochondria and the Golgi apparatus to provide a framework for their classification. The image and structural data obtained by this process was used to guide the development of an image processing pipeline for the semi-automated segmentation of specific classes of these organelles on the path to developing more advanced automated processes.
Chapter 1 provides an overview of biology of pancreatic beta cells and the process of insulin secretion as well as electron tomography and the motivation for the development of automated segmentation processes.
Chapter 2 describes the methods used to prepare and analyse these data sets. A number of cellular tomograms of insulin-secreting pancreatic beta cells recorded at high (i.e. 4-5 nm) resolution were used as primary ‘proof-of-concept’ datasets for this project. To control the number of datasets (i.e. organelle sub-volumes) for the project, three key organelles of insulin secretion were selected; the Golgi apparatus (GA), mitochondria (MC) and insulin granules (IG).
Chapter 3 introduces and describes the proposed segmentation pipeline which is referred to as the ‘cellular tomography segmentation’ (CTS) workflow. It also introduces a scoring system that is based on mesh surface area (MSA) of an organelle’s 3D model and provides a useful, quantitative comparison for assessing the quality of various segmentation approaches, compared with the results obtained by manual tracing. Overall this chapter covers significant computational considerations for the development of segmentation algorithms for electron tomography.
Chapter 4 introduces the concept of the categorisation of sub-cellular organelles according to their image properties both to provide a basis for morphological classification of organellar subtypes and to enable improved image segmentation of each of these subclasses. The performances of tracing tools are quantitatively compared and conclusions on best performance drawn in this chapter.
Concurrent with the research described here, a new filter for automated edge detection-based processing of 3D volumetric image data was developed in the Hankamer Lab at the University of Queensland. The 3D Bilateral Edge detector (3D BLE) (Ali et al., 2012) yielded their first successful results in automatically segmenting organelles in high resolution electron tomograms. In Chapter 5 the 3D BLE filter was used to segment the selected data sets for comparison with semi-automated processes based on the cellular tomography segmentation’ (CTS) workflow. The segmentation results obtained using the best detected settings identified for each organelle sub-volume were compared to those obtained using the semi-automated CTS workflow, as outlined in Chapters 3 and 4. This comparison suggested that for the datasets and conditions analysed, the CTS workflow was superior in performance, both in terms of quantitative and qualitative comparison. In terms of quality it proved comparable to manual tracing but also better than the best detected 3D BLE settings.
Overall, this newly developed semi-automated CTS workflow and the image categorisation technique enable improved rates of segmentation of sub-cellular compartments. It also enables rapid, quantitative comparison of the morphology and function of three key organelles of insulin secretion of non-stimulated pancreatic beta cells. It also offered sets of scoring objectives for different organelle sub-groups to expedite the process of optimising method settings not currently afforded by any other technique.