Detección supervisada de anomalías en tomografías torácicas mediante autoencoders y clustering

Abstract

CONDITION FOR PUBLICATION OF PROJECT. project addresses the early detection of anomalies in chest computed tomography within the framework of preventive healthcare at the university level, aiming to support clinical diagnosis through supervised artificial intelligence. The hypothesis states that a deep learning model, trained with preprocessed and balanced data, can identify pulmonary abnormalities, contributing to institutional epidemiological surveillance and aligning with SDG 3: Good Health and Well-being. For its development, images were collected and filtered from public databases such as NIH ChestX-ray8, incorporating additional tuberculosis and pneumonia datasets. Preprocessing techniques such as thoracic segmentation, CLAHE, and normalization were applied. Ten models between machine learning and deep learning were trained and their performance was compared using AUC-ROC, accuracy, and confusion matrices. The results showed that DenseNet121, trained with a combined dataset of original and segmented images, achieved the best performance (accuracy: 0.94, AUC: 0.91). Moreover, segmentation improved predictions in most classes, although it introduced limitations in specific cases such as emphysema. It is concluded that supervised AI enhances automated diagnosis, and that preprocessing quality and class balance are key factors in optimizing outcomes. Keywords: artificial intelligence, chest tomography, anomaly detection, supervised learning