Segmentação Automática de Cromossomas em Imagens Microscópicas de Cariótipos ; Automatic Segmentation of Chromosomes in Microscopic Images of Karyotypes

Trabalho de Projeto do Mestrado Integrado em Engenharia Biomédica apresentado à Faculdade de Ciências e Tecnologia ; Since the study of the karyotype entails the examination of a person's unique set of chromosomes, conventional cytogenetics plays a significant part in the diagnosis and prognosis of several types of cancer as well as in the surveillance of residual diseases.The automated karyotype analysis seeks to create a karyogram that is properly annotated and prepared for the expert's review. The chromosomal detection method can be automated to increase speed and accuracy of karyotyping. Due to the difficulties in segmenting chromosomal clusters, chromosome segmentation has not yet been widely used in clinical settings. Additionally, a barrier to automating this procedure is the lack of clinical datasets or photorealistic synthetic datasets.This research proposes an algorithm to perform the segmentation of chromosomes in microphotographies of chromosomes exhibiting a G-band pattern. By means of a "Cut, Paste, and Learn" procedure a method to generate synthetic images with great morphological variability is described, while still being based on actual cell structures of the cytogenetic clinic. The algorithm YOLOv5 was used to detect chromosomes with a G-band pattern in real images.A dataset of real cell structures was produced during the “Cut” phase, including 115896 chromosomes, 180 nucleoli and 6024 noisy objects. Using a proposed blending technique for smoothing the overlapping of cell structures, 10795 synthetic images were created during the “Paste” phase. During the “Learn” phase a value of mAP$@0.5$ equal to 0.989 was obtained for the validation group, which included 1080 synthetic cell images. The LCG-FMUC dataset, which consists of 171 microphotographs of cells with prophase or metaphase chromosomes, was used to test the YOLOv5 model. Overall, from the 7861 chromosomes that were contained in the experimental dataset, 7708 chromosomes were correctly segmented which accounts for 98.05% of success in ...