IMF-DeepMarble: Deep learning for precise marbling grade prediction and intramuscular fat content analysis in pork quality assessment

• Published in: Journal of agriculture and food research Vol. 23; p. 102251
• Main Authors: Yang, Xidi, Zhu, Liangyu, Jiang, Wenyu, Cheng, Jipeng, Wang, Songlin, Pu, Xinyi, Yang, Yiting, Gan, Mailin, Shen, Linyuan, Zhu, Li
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.10.2025; Elsevier
• Subjects: Intramuscular fat content; Marbling; ResNet; SE-Attention; Intramuscular fat content; Marbling; SE-Attention; ResNet
• ISSN: 2666-1543; 2666-1543
• DOI: 10.1016/j.jafr.2025.102251

The degree of marbling in pork is a key indicator of its quality, with richer marbling indicating more tender and palatable meat. The formation of marbling is influenced by factors such as genetics, management practices, and growth cycles, and it can be measured through Intramuscular Fat Content (IMF%). Higher IMF% generally indicates superior meat quality, directly affecting the market value of pork and consumer preferences. This study presents a new deep learning model called IMF-DeepMarble, which integrates Convolutional Neural Networks (CNN) and Transformer technology to assess IMF content using images of pork longissimus dorsi slices. We utilized a high-resolution camera to capture images of the pork slices and employed the Soxhlet extraction method to obtain accurate IMF values. Experimental results demonstrated that integrating attention mechanisms into Convolutional Neural Networks (CNNs) significantly enhanced pork quality prediction. In intramuscular fat (IMF) content prediction, ResNet + SE and ResNet + CBAM models achieved the highest Pearson correlation coefficient (PCC) of 0.42, outperforming baseline ResNet (PCC = 0.12) and other architectures (DenseNet/EfficientNet showed poor performance). For marbling grade prediction, the optimized VGG + SE model achieved the highest accuracy (0.90) and ROC-AUC (0.95), surpassing the baseline VGG (0.81/0.88) and ResNet (0.85/0.94). Notably, the generalizability of IMF-DeepMarble to other meat types (e.g., beef, lamb) remains unvalidated, and its performance is dependent on image quality (e.g., lighting, slicing consistency). Future work should involve expanding datasets across species and developing robust preprocessing pipelines to mitigate environmental variability, thereby enhancing the practical applicability of IMF-DeepMarble in diverse meat quality assessment scenarios. [Display omitted] •IMF-DeepMarble fuses CNN-Transformer for 0.90 PCC marbling/IMF assessment.•Image-based IMF analysis reduces lab dependency.•SE-ResNet improves marbling PCC by 18.4 % via attention mechanisms.•Preprocessing pipeline enhances pork image contrast and noise reduction.•IMF-DeepMarble enables real-time meat quality control across supply chains.