A pilot study on automated multi-task deep learning framework for pregnancy diagnostic and gestational age estimation in goats based on ultrasonography

Background: This pilot study aims to develop an automated, multi-task deep learning system as a proof of concept for both the classification of pregnancy status and gestational day prediction in Korean crossbred goats.Methods: Two 1-year-old goats were followed longitudinally until mid to late gestation.A complete data set containing 11,092 high-resolution image frames (6,647 pregnant and 4,445 non-pregnant) was created from the video sequences.The dataset was segmented at the frame level with 80% allocated to the training partition (n = 8,875), 10% to the validation partition (n = 1,110), and 10% to the test partition (n = 1,107).Multi-task Four convolutional neural networks (ResNet18, EfficientNet-B0, MobileNetV3-Large, and ConvNeXt-Tiny) were trained following a multi-task learning strategy with the AdamW optimizer, OneCycle scheduling, and a hybrid Binary Cross-Entropy and Huber loss function.Results: Accuracy was achieved for the four architectures, with more than 99.5%.The MobileNetV3-Large model performed the best in terms of frame-level classification accuracy, with 100% sensitivity and specificity.For the regression-based task for predicting gestational days, EfficientNet-B0 got the best performance with a framelevel Mean Absolute Error (MAE) of 1.3 days.The video-level MAE of 0.882 days and the competitive video MAE of 1.204 days were the outcomes for the aggregated video sequences using ConvNeXt-Tiny and EfficientNet-B0, respectively.Conclusions: The present pilot study lays the foundation for the methodological feasibility of small optimized deep learning models for accurate monitoring of fetal development and for the automation of reproductive diagnostics.