Answers are grounded in this paper and cite their sources.
Despite its importance for chicken meat quality and nutrition, fat content is traditionally quantified using methods that are laborious, time-consuming, and inaccurate. To address these challenges, an integrated workflow for chicken fat quantification based on computed tomography (CT) was proposed. A total of 200 chicken thigh samples (4,950 CT slices) were used, with reference fat content manually measured for 50 samples. A high-precision consumption enhancement model was developed using a conditional Generative Adversarial Network (cGAN) to improve the quality of CT images. U-Net was then employed as the backbone segmentation network, and a Transformer-based context bridge was introduced to capture multi-scale contextual relationships. Multiple CT slices from the same sample were analyzed to calculate the fat pixel ratio, which served as a key input feature for a machine learning model to predict fat content. Experimental results demonstrate that our framework significantly outperforms competing models, achieving 90.36% IoU, 94.70% Dice, 94.60% Precision, and 94.88% Recall in segmentation. For fat content prediction, it attained a high correlation (r=0.917) and low root mean square error (RMSE=0.0041) with ground truth. This end-to-end framework integrates GAN-based image enhancement, transformer-driven segmentation, and multi-slice feature aggregation, providing a novel and effective solution for accurate and efficient food component quantification.
No takes yet. Share an insight, caveat, or question.
Wei et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: