Content-based comparison of traditional methods and convolutional neural networks in medical images

In recent years, with the development of computer technologies and the rapid increase in internet usage, image retrieval systems have gained importance. In this thesis, content based image retrieval systems which aim to increase the speed of image access and decrease the storage space requirement are examined. Content-based image retrieval systems (CBIR) are widely used in many areas, including health. Today, medical imaging systems are widely used in the diagnosis of many diseases. Different models such as ultrasound, tomography, x-ray, magnetic resonance imaging are widely preferred by experts. Although these models have different working principles, they are based on the acquisition of images of identified areas of the patient from different angles. As a result, the number of medical images increases day by day. CBIR systems can be used to access medical images quickly and accurately when needed. In this thesis, it is aimed to design CBIR system with two different methods for medical images. In the first design, color, texture and shape features of medical images were extracted. Features were compared using simple metrics to measure similarity between images. In the second design, instead of feature extraction, deep learning techniques was followed. The results of both designs were presented and reviewed in the thesis. It will be easier for staff and doctors working in health units to quickly examine similar cases before medical diagnosis with the proposed system.