Segmentación de lesiones en imágenes de colposcopia usando redes neuronales artificiales

Abstract

CONDITION FOR PUBLICATION OF PROJECT. Cervical cancer is the fourth leading cause of mortality in women, according to the Global Cancer Observatory and GLOBOCAN 2022 worldwide. Early detection through techniques such as colposcopy is crucial to reducing mortality. The project aims to develop an algorithm for automatic lesion segmentation in colposcopic images using U-Net-based artificial neural networks. This approach stems from the need to automate medical image analysis to segment cervical lesions and reduce errors caused by physicians' subjective interpretation. The project used a database of 200 colposcopy images from the Kaggle platform, in which regions of interest (ROIs) were manually segmented with the collaboration of trained medical personnel to generate binary masks necessary to train the model. Data augmentation techniques were employed, and performance was evaluated using the Dice coefficient and Intersection over Union (IoU) metrics. The results showed that the model achieved a Dice coefficient of 84% on the validation set and the IoU achieved 75%. The U-Net model proved effective in segmenting lesions in colposcopy images, with adequate validation performance. The model could be used in an assistance tool to assist physicians in segmenting cervical cancer lesions. Keywords: Artificial Intelligence, Colposcopy, Cervix, U-Net, ROI.