Noninvasive estimation of porcine intramuscular fat from ultrasonography using deep learning

Intramuscular fat is a key determinant of pork quality. This study evaluated whether deep learning could be applied to ultrasonographic images to noninvasively estimate porcine intramuscular fat. We analyzed 9,409 single-channel brightness mode images acquired using a 3.5‑MHz system targeting the longissimus dorsi muscle and subcostal region. Two models were trained and compared using cross‑validation—that is, Xception, a convolutional neural network, and Vision Transformer‑Huge (ViT‑H). The Xception model employs depthwise separable convolutions to efficiently capture fine-grained local textures in ultrasound images. Conversely, the ViT‑H model tokenizes images into fixed‑size patches and uses multi‑head self‑attention to model long‑range dependencies and global context, providing a complementary inductive bias to convolutional neural networks. We also examined the effects of data augmentation and denoising. Model performance was evaluated using the root mean squared error (primary metric), mean absolute error, and coefficient of determination metrics. The best configuration—that is, the ViT‑H model with augmentation and denoising—achieved a root mean squared error of 0.3559%, mean absolute error of 0.2172%, and coefficient of determination of 0.8459 on three independent test set, outperforming the baselines. These findings indicate that ultrasonography with deep learning can enable non-invasive estimation of porcine intramuscular fat, which suggests the potential for on-farm and near-real-time monitoring.