Machine Learning Algorithm in the Classification of frontal view Chest x-ray as “normal” or “pathological”

• Autor(a) principal: Cidade de Moura, Ivone
• Data de Publicação: 2023
• Outros Autores: Mesquita, Luís, Teresa Ribeiro, Ricardo
• Tipo de documento: Artigo
• Idioma: por
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: https://doi.org/10.46885/roentgen.v4i2.114
• topic: artificial intelligence; computer-aided detection; machine learning classifiers; chest x-ray; principal component analysis; rank; inteligência artificial; computer-aided detection; classificadores de aprendizagem automática; radiografia tórax; principal component analysis; rank

The growing demand for radiological exams puts pressure on imaging services that face challenges due to the shortage of radiologists, requiring faster interpretation with a higher potential for error. In parallel, radiomics and artificial intelligence (AI) techniques have proven to be important tools in the field of radiology, revolutionizing clinical practice itself. Using these techniques, this paper developed a pathology detection classifier for chest radiographs, from the public database CestXray14, in order to highlight the crucial role that understanding radiomics techniques and AI play in the radiographer's profession.  We analysed 1662 chest radiographs (50% with pathology) and applied two strategies for selecting 5 radiomics features: (i) principal component analysis and (ii) information gain ratio, using the Orange software. With the PCA method, with reduction to 5 components and 73% of variance explained, the best classifier was the Neural Network, with 0.987 of Area Under the Curve (AUC). In Information Gain Ratio also the Neural Network was the best classifier with 0.972 AUC, in which a sensitivity of 97.8%, specificity of 92.9% and accuracy of 93% were found.  By using AI techniques and taking advantage of a large dataset, our study demonstrates the feasibility of using automatic classifiers to aid in the interpretation of chest radiographs, indicating their potential as a valuable tool in screening, prioritizing exams, and optimizing workflow in radiology departments.