Point-of-care diagnostic imaging is crucial in various veterinary and agricultural settings. In sheep, pregnancy diagnosis by ultrasound is routinely undertaken to determine pregnancy status, fetal number, age and even health. However, diagnostic accuracy remains challenging, even for highly experienced scanners. Deep learning holds potential to address such challenges and improve accuracy and efficiency of ultrasound imagery interpretation. This project aimed to evaluate the ability of a deep learning model to identify pregnant and non-pregnant ewes from ultrasound imagery.
Transcutaneous ultrasound (OviScan 6, BCF Australia, Mitcham VIC) imagery was recorded from 937 ewes, 44 days post 5-week joining. From these recordings, images were created every 1/12 second and labelled with a diagnosis class (pregnant or non-pregnant), determined by a trained scanner. To achieve class balance, 1,838 images from 62 pregnant ewes and 1,826 images from 48 non-pregnant ewes were randomly split between training (70%), validation (20%), and testing (10%) datasets. A binary classification Convolutional Neural Network (1) was trained on the training and validation datasets, using binary cross-entropy loss. The architecture included two convolutional layers and a fully connected layer, using a batch size of 32. The trained model was then evaluated on the testing dataset, generating a confusion matrix for precision and recall assessment. Grad-CAM (2) was used to visualise pixels of importance. Testing accuracy was 98.92%. Accuracy was 98.19% for pregnant ewes (1.00 precision; 0.99 F1-score; 0.98 recall) and 100% for empty ewes (0.97 precision; 0.99 F1-score; 1.00 recall).
This study shows deep learning is highly accurate at identifying pregnant and non-pregnant ewes, giving confidence for future detailed classifications including fetal number, age, health, and placentome types. This model demonstrated its potential to diagnose pregnancy quickly and accurately for on-farm sheep management, as well as to classify ultrasounds of other species and humans for research, treatment, and animal management.