RAT MESENCHYMAL STEM CELL SEGMENTATION ON TIME-LAPSE FLUORESCENCE MICROSCOPY VIDEO WITH TWO-DIMENSIONAL BANDPASS FILTERS

• Main Author: Firdaus, Lulu
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/75282
• Institution: Institut Teknologi Bandung

Rat mesenchymal stem cells (MSC) is essential in the development of regenerative medicine. Cell movements and shape changes as they differentiate are the key aspects analyzed in this field. As live cell imaging tools advance, the amount of data generated in this field also grows. Thus, research in developing faster and more precise segmentation and tracking methods became popular. Cell Tracking Challenge (CTC) was initiated in International Symposium of Biomedical Imaging (ISBI) 2013 to compare each method quantitively. In CTC 2020, bandpass segmentation ranked second for rat MSC cell segmentation. Analysis of this method revealed its potential to increase its segmentation performance through parameter optimization. This research will discuss the parameter optimization of two Gaussian filters and a tophat filter involved in bandpass segmentation. Optimization techniques used in Fluo-C2DL-MSC dataset in CTC are grid search and perceptron model, aimed for different parameters. Through this optimization, we could determine the characteristics of bandpass filters with the best segmentation results and the cells they managed to segment well. Results concluded by this study are good segmentation images were produced by bandpass filters, which simultaneously possess a Mexican Hat shape in their spatial domain. To create these filters, the ratio between two Gaussian standard deviations must be preserved around ?0.28 to ?0.99. The size of tophat filters which supported these results is high, ranged between 300 to 1000 pixels. This parameter optimization acquired segmentation results of Jaccard Index (JI) 0.708±0.002, 0.728±0.002, and 0.739±0.004 for 4 training images (37 cells), 4 validation images (36 cells), and 40 testing images. In three image groups, optimized bandpass segmentation managed to segment roughly 66.67% of cells with JI score higher than each group mean value and 33.34% cells scoring lower than the mean. Cells with lower JI score tends to be thin and long, horizontal/vertical, sticking to the upper side of the image, their area and major axis below their 3rd decile, and their minor axis under its 2nd decile.