Identifying pregnancy in cows using ovarian ultrasound images and convolutional neural networks - a proof-of-concept study

•The dataset included 510 images from the individual functional ovaries of cows 30 days after insemination.•CNN architectures ResNet50, ResNeXt50, InceptionResNetV2, and DenseNet-121 were evaluated, with two datasets varying in image quality.•Higher accuracy of 78.8% was observed for DenseNet-121 architecture.•Image quality affects CNN performance differently according to the architecture.•CNNs are a promising support tool for pregnancy diagnosis in cows. Visual analysis of ovarian structures by ultrasound in cows is a relevant support tool for improving reproductive performance in livestock. However, with human visual analysis, subjectiveness is a limiting factor, and computer vision technologies are a way to overcome this. We thus aimed to evaluate the use of convolutional neural networks (CNNs) for identifying pregnancy in cows using ovarian ultrasound images obtained 30 days after artificial insemination. The dataset consisted of 510 images from individual independent functional ovaries of 238 pregnant and 272 non-pregnant Nellore cows. All the images were collected from the same commercial farm. To evaluate the dependency of CNN performance on image quality, the images were classified by two independent veterinarians with significant experience in ultrasound evaluation as good, regular, or bad for visual identification of ovarian structures. Four CNN architectures were evaluated, namely ResNet50, ResNeXt50, InceptionResNetV2, and DenseNet121. All the CNNs were evaluated using the complete dataset (ALL; n = 510) and a subset of good and regular images (GR; n = 462). Ten evaluations were conducted for each architecture and dataset combination. In each run, the datasets were randomly divided into 70 % for training and 30 % for testing according to the holdout method. A regularization-based method was used during the training phase using dropout, data augmentation, and a dynamic learning rate. The model’s performance was evaluated in terms of accuracy, precision, sensitivity, and specificity, which ranged from 46.5 to 78.8, 34.4–80.0, 28.3–75.4, and 11.2–82.2 %, respectively. The DenseNet121 architecture performed better based on higher accuracy, sensitivity, and specificity, and both lower SD and variation due to dataset. ResNet50 and DenseNet121 models performed better when using the GR dataset. InceptionResNetV2 had a more precise architecture when using the ALL dataset but had lower performance for the GR dataset. We concluded that CNNs are a promisin