Deep learning–based automated measurements of the scrotal circumference of Norwegian Red bulls from 3D images

Abstract

The main aim of this study was to create an automated method for the measurement of the scrotal circumference (SC) of Norwegian Red bulls using 3D images of the scrotum based on convolutional neural networks. The study population was bull calves recruited for performance testing before the selection of bulls for semen production in the breeding program. Bulls were measured at four different time points: upon arrival in quarantine (Q) and thereafter at approximately 6, 9 and 12 months of age. Both 3D images and manual SC measurements were performed at all time points. In our approach, SC could be calculated without direct contact with the bull, using only 3D images and a simple, user–friendly application into which mentioned images are uploaded. The results show that SC measurements obtained using semantic segmentation are comparable with manual measurements. The mean prediction error was significantly different between age groups Q, 6, 9 and 12, and it was -3.07 cm, -3.02 cm, -1.79 cm and -1.11 cm, respectively. The results show a significant difference in the measurement error of the SC based on the quality of the images. Images were categorised into three quality groups. For good prediction accuracy, we recommend capturing 3D images of quality 2 – full circle from individuals older than 6 months.

Keywords

• Convolutional Neural Networks
• Semantic segmentation
• U–Net
• 3D imaging
• Scrotal circumference
• Bull