Ground truth determination for segmentation of tomographic volumes using interpolation

Dissertação para obtenção do Grau de Mestre em Engenharia Biomédica

Optical projection tomographic microscopy allows for a 3D analysis of individual cells, making it possible to study its morphology. The 3D imagining technique used in this thesis uses white light excitation to image stained cells, and is referred to as single-cell optical computed tomography (cell CT). Studies have shown that morphological characteristics of the cell and its nucleus are deterministic in cancer diagnoses. For a more complete and accurate analysis of these characteristics, a fully-automated analysis of the single-cell 3D tomographic images can be done. The first step is segmenting the image into the different cell components. To assess how accurate the segmentation is, there is a need to determine ground truth of the automated segmentation. This dissertation intends to expose a method of obtaining ground truth for 3D segmentation of single cells. This was achieved by developing a software in CSharp. The software allows the user to input a visual segmentation of each 2D slice of a 3D volume by using a pen to trace the visually identified boundary of a cell component on a tablet. With this information, the software creates a segmentation of a 3D tomographic image that is a result of human visual segmentation. To increase the speed of this process, interpolation algorithms can be used. Since it is very time consuming to draw on every slice the user can skip slices. Interpolation algorithms are used to interpolate on the skipped slices. Five different interpolation algorithms were written: Linear Interpolation, Gaussian splat, Marching Cubes, Unorganized Points, and Delaunay Triangulation. To evaluate the performance of each interpolation algorithm the following evaluation metrics were used: Jaccard Similarity, Dice Coefficient, Specificity and Sensitivity.After evaluating each interpolation method we concluded that linear interpolation was the most accurate interpolation method, producing the best segmented volume for a faster ground truth determination method.