Divisible cell-segmentation: uma nova abordagem fine-tuning adaptativa para segmentação em tomografias computadorizadas por meio de células computacionais sintéticas

Abstract

According to the World Health Organization (WHO), various pathologies are factors causing millions of deaths per year worldwide, such as Stroke and pulmonary diseases. With the aim of optimizing and facilitating the process of localization and segmentation of regions in medical images, computer-aided diagnosis (CAD) systems are increasingly being used, increasing the safety of diagnoses and the performance of specialists. The objective of this thesis is to present a new approach for segmentation and automatic initialization in computed tomography (CT) images of hemorrhagic stroke and pulmonary CTs. The study proposes a new, effective, fully automatic, and adaptive method for segmentation in medical images, called Divisible Cell-Segmentation. The proposed method is an effective new approach based on automatic initialization for detection and segmentation of the region of interest, where synthetic computational cells capable of dividing (multiplying) with the aim of finding the edges of the object of interest, effectively and accurately segment the contour regions in CT medical images. The synthetic computational cell division process is based on an analogy to the biological process of cytokinesis, the process of cell division where cells multiply in the human body. The experiments were successful, achieving high effectiveness in segmenting the regions in different medical images, obtaining 99.80% accuracy for segmentation in hemorrhagic stroke, with 99.81% specificity (ESP), demonstrating effectiveness in segmentation in different brain CTs. For pulmonary images, the proposed method achieved 99.01% accuracy, being able to identify different lung regions and their structures, as well as their different shapes caused by the breathing process in the acquisition of exams. As in hemorrhagic stroke, segmentation in lung images obtained excellent results with 99.28% specificity, demonstrating the robustness of the new method with an average segmentation time of just over 1 second for CT images. To validate the experiments, the results were compared to different methods found in the state of the art, thus surpassing renowned computational segmentation models in the literature; non-automatic, semi-automatic, and automatic, equaling to consolidated state-of-the-art methods in segmentation time as well as in segmentation effectiveness.