IQAEvolNet: A Novel Unsupervised Evolutionary Image Enhancement Algorithm on Chest X-Ray Scans

• Ano de defesa: 2024
• Autor(a) principal: Vasconcelos Filho, Carlos Alfredo Cordeiro de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Palavras-chave em Português: Aprimoramento de imagens; Rede neural convolucional; Aprendizado não supervisionado; Algoritmo genético; Raio-x; Imagem médica; Image enhancement; Convolutional Neural Network; Unsupervised learning; Evolutionary algorithms; X-ray; Medical image; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufc.br/handle/riufc/78756

X-ray images are widely used in the medical field due to their low cost and non-invasive nature, but they suffer from noise problems related to equipment or environmental factors. There are multiple solutions in the literature to combat this problem, with the main one being the use of non-learning image processing algorithms for X-ray image enhancement. In other areas of application, such as low-light and underwater images, there is an extensive use of artificial intelligence models for image enhancement tasks, but the training of artificial intelligence models for medical image enhancement encounters significant challenges. In supervised learning, obtaining a dataset with authentic noisy images and their manually enhanced counterparts as labels is imperative. When dealing with medical images it can be difficult to have access to high-quality/low-quality pairs because of the restrictive context where these images are taken. To deal with this problem, this paper introduces an innovative approach to unsupervised learning for chest x-ray image enhancement. The suggested approach begins with the pre-training of a model using multiple image enhancement algorithms as reference to establish an initial set of solutions. Following this, an evolutionary algorithm is employed to refine these initial solutions. It incorporates two image enhancement metrics, Entropy and the Natural Image Quality Evaluator(NIQE), along with Structural Similarity Index as fitness indicators. We tested our method in a Chest X-ray dataset and our findings demonstrate that our method achieved a better NIQE, 4.05 compared to 4.24, and a faster processing time, 2.95 milliseconds compared to 0.195 seconds, in relation to the state-of-the-art algorithm with the best NIQE and entropy. We showed that our algorithm outperforms state-of-the-art algorithms in NIQE and processing time.