Detecção de ovócitos em imagens histológicas utilizando Redes Neurais Convolucionais

Abstract

Oocytes are complex germ cells that present structural and functional changes throughout maturational development in fish. Therefore, it is of significant importance to develop approaches that perform the automatic detection of oocytes, since distinguishing them in histological images is a time-consuming and error-prone procedure, as it requires the visual and subjective interpretation of the specialist. Therefore, this work proposes a method capable of automatically detecting them. For this, an image database containing 5,680 oocytes was used, which was divided into training, testing and validation. Data augmentation techniques were employed to make the models more robust. To perform the detection, the YOLOv3, YOLOv4 and YOLOv5 architectures were applied to the image base. The results obtained were promising, highlighting that the YOLOv5m and YOLOv5l models achieved the best performances in oocyte detection, when evaluated using the Precision, recall and mAP@.5 and mAP@.95 metrics