Redes neurais convolucionais aplicadas a imagens de ultrassonografia de ovário para identificar a gestação em vacas nelore.

Abstract

Visual analysis of ovarian structures by ultrasound in cows are a relevant support tool for improving reproductive performance in livestock. However, by human visual analysis the subjectiveness is a limiting factor and computer vision technologies are a good way to overcome it. We aimed to evaluate the use of convolutional neural networks (CNNs) for identifying pregnancy in cows using ovarian ultrasound images obtained at 30 days after artificial insemination. A dataset with 510 images from individual functional ovaries of 238 pregnant and 272 non-pregnant Nellore cows was used. All images were collected in a same commercial farm. To evaluate dependency of CNNs performance on image quality, the images were classified by two independent veterinarians with significant experience in ultrasound evaluation as: good, regular, and bad for visual identification of ovarian structures. Five CNNs architectures were evaluated: ResNet50, ResNeXt50, InceptionResNetV2, DenseNet-121 and SimpleCNN. All CNNs were evaluated both using the complete dataset (ALL; n=510) and a subset of good and regular images (GR; n=462). For each architecture and dataset combination, ten evaluations were performed. In each run, the data sets were randomly divided into 70% for training and 30% for testing according to holdout method. A regularization-based method via dropout, data augmentation and a dynamic learning rate during the training phase was used. The model’s performance was evaluated by accuracy, precision, sensitivity, and specificity, which ranged from 0.648 to 0.788, 0.344 to 0.800, 0.283 to 0.754, and 0.112 to 0.822, respectively. The models ResNet50, DenseNet-121 e SimpleCNN performed better when using the GR dataset. The DenseNet121 architecture performed better based on accuracy, sensitivity, specificity and with lower variation among datasets. The InceptionResNetV2 the more precise architecture using ALL, however with lower performance for GR dataset. We concluded that CNNs are capable to identify cow’s pregnancy status in ovarian images collected at 30 days after artificial insemination. The effect of image quality on performance is dependent on CNNs architecture.