Body condition scoring is a simple method to estimate the energy supply of dairy cattle. Our study aims to investigate the accuracy with which supervised machine learning, specifically a deep convolutional neural network (CNN), can be used to retrieve body condition score (BCS) classes estimated by an expert. We recorded images of animals’ rumps in three large-scale farms using a simple action camera. The images were annotated with classes and three different-sized bounding boxes by an expert. A CNN pretrained model was fine-tuned on 12 and 3 BCS classes. Training in 12 classes with a 0 error range, the Cohen’s kappa value yielded minimal agreement between the model predictions and ground truth. Allowing an error range of 0.25, we obtained minimum or weak agreement. With an error range of 0.5, we had strong or almost perfect agreement. The kappa values for the approach trained on three classes show that we can classify all animals into BCS categories with at least moderate agreement. Furthermore, CNNs trained on 3 BCS classes showed a remarkably higher proportion of strong agreement than those trained in 12 classes. The prediction precision when training with various annotation region sizes showed no meaningful differences. The weights of our trained CNNs are freely available, supporting similar works.

Statistical learning algorithms strongly rely on an oversimplified assumption for optimal performance, that is, source (training) and target (testing) data are independent and identically distributed. Variation in human tissue, physician labeling and physical imaging parameters (PIPs) in the generative process, yield medical image datasets with statistics that render this central assumption false. When deploying models, new examples are often out of distribution with respect to training data, thus, training robust dependable and predictive models is still a challenge in medical imaging with significant accuracy drops common for deployed models. This statistical variation between training and testing data is referred to as domain shift (DS).To the best of our knowledge we provide the first empirical evidence that variation in PIPs between test and train medical image datasets is a significant driver of DS and model generalization error is correlated with this variance. We show significant covariate shift occurs due to a selection bias in sampling from a small area of PIP space for both inter and intra-hospital regimes. In order to show this, we control for population shift, prevalence shift, data selection biases and annotation biases to investigate the sole effect of the physical generation process on model generalization for a proxy task of age group estimation on a combined 44 k image mammogram dataset collected from five hospitals.We hypothesize that training data should be sampled evenly from PIP space to produce the most robust models and hope this study provides motivation to retain medical image generation metadata that is almost always discarded or redacted in open source datasets. This metadata measured with standard international units can provide a universal regularizing anchor between distributions generated across the world for all current and future imaging modalities.

Body condition scoring is a simple method to estimate the energy supply of dairy cattle. Our study aimed to investigate the accuracy with which supervised machine learning, a deep convolutional neural network, can be used to retrieve body condition score (BCS) classes estimated by an expert. Using a simple action camera, we recorded images of animals' rumps in three large-scale farms. The images were annotated with three different-sized boxes by an expert. A Faster-RCNN pre-trained model was trained on 12 and 3 BCS classes. Training in 12 classes, with a 0 error range, the Cohen's kappa value yielded minimal agreement. Allowing an error range of 0.25, we obtained a minimum or week agreement. With an error range of 0.5, we had strong or almost perfect agreements. The kappa values of the approach trained on 3 classes show that we can classify all animals into BCS categories with at least moderate agreement. Furthermore, CNNs trained in 3 BCS classes show a remarkably higher proportion of strong agreement than those trained in 12 classes. The prediction precision based on training with various annotation regions showed no meaningful differences.

Leveraging recent advances in computational modeling of proteins with AlphaFold2 (AF2) we provide a complete curated data set of all single mutations from each of the 7 main SARS-CoV-2 lineages spike protein receptor binding domain (RBD) resulting in 3819X7 = 26733 PDB structures. We visualize the generated structures and show that AF2 pLDDT values are correlated with state-of-the-art disorder approximations, implying some internal protein dynamics are also captured by the model. Joint increasing mutational coverage of both structural and phenotype data coupled with advances in machine learning can be leveraged to accelerate virology research, specifically future variant prediction. We hope this data release can offer assistance into further understanding of the local and global mutational landscape of SARS-CoV-2 as well as provide insight into the biological understanding that 3D structure acts as a bridge between protein genotype and phenotype.