EIMFS: Estimating intramuscular fat in sheep using a three-stage convolutional neural network based on ultrasound images

Abstract

Intramuscular Fat (IMF) is a key factor in meat quality, significantly affecting the tenderness, juiciness, and flavor of mutton. The non-invasive approach for estimating live sheep IMF content is essential for sheep breeding. In this study, we proposed a three-stage convolutional neural network (CNN) called EIMFS to estimate IMF in sheep based on ultrasound images. Our proposed method first segments loin areas from captured images and generates segmentation masks. These masks are then concatenated with the original color and grayscale ultrasound images, respectively. Loin areas are also estimated from the segmentation masks. Through a multi-branch, IMF estimation features are extracted from masked color and grayscale images and are fused with linearly mapped loin area estimations. Finally, IMF values are estimated based on the fused multi-dimensional features. The proposed model was trained and tested on a manually annotated sheep backfat ultrasound image dataset. The results showed that the mean absolute percentage error (MAPE) of the IMF estimation was 7.25%, and the intraclass correlation coefficient (ICC) between EIMFS and the Soxhlet extract method was 0.905. Compared to existing deep learning approaches, the proposed approach significantly lowered IMF estimation error, and can enable real-time estimation and long-term monitoring of IMF content in sheep.