Genetic algorithms for hyperparameter tuning of a DC-UNet Model for medical image segmentation

Computer vision is a branch of artificial intelligence that enables computers to extract information from images and perform tasks such as image segmentation, which involves identifying multiple elements as image regions. Then, the application of image segmentation in the medical area is used to assist physicians in disease diagnosis from patients based on visual information, and they are required to have the best possible accuracy. In this work, a computer vision model called dual channel U-Net (DC-UNet) is used for medical image segmentation. Specifically, we focus on the area of polyp detection which are lesions that can vary in their size from a few millimeters to several centimeters, and the importance of this application relies on the early identification for colorectal cancer prevention. One of the most challenging public datasets in this field called CVC-ClinicDB was employed to train the segmentation model. These medical images correspond to colonoscopy video frames, whose ground truth images consist of a fully annotated binary segmentation between polyp and background. Furthermore, to increase the performance of the DC-UNet model on this challenging dataset, we propose a genetic algorithm that finds the optimal hyperparameter combination for this specific application. Finally, we use different genetic configurations to study the performance of some state of the art gradient-based optimizers regarding this task.

• Автор: Román Eras, Krishna Gautama (author)
• Формат: bachelorThesis
• Мова: eng
• Опубліковано: 2023
• Предмети: Algoritmos genéticos; Visión por computador; Genetic algorithms; Computer vision; Medical image segmentation
• Онлайн доступ: http://repositorio.yachaytech.edu.ec/handle/123456789/678