Deep learning-based classification and internal region stratification of wooden breast in broiler by using ultrasound imaging

Currently, palpation remains the predominant method for classifying wooden breast (WB). This approach requires considerable labor and time resources, and it fails to precisely characterize the complex internal structural distribution of the disease and lacks a rational utilization plan for WB-affected breast fillets. Thus, the scientific stratification and classification of WB must be investigated. This study aims to characterize WB severity using ultrasound-derived internal spatial information, combined with ImageJ threshold binarization and scale calibration to quantify the spatial extent of pathological features. Herein, chicken breast fillets from Arbor Acres broilers were collected (n = 240, males, 42 days old) and categorized into four categories: normal (NORM, n = 60), mild (MILD, n = 60), moderate (MOD, n = 60), and severe (SEV, n = 60) conditions. WB samples were classified via ultrasound scanning and deep learning (DL). MobileNetV3, ResNet18, and AlexNet achieved training accuracies of 99.50%, 96.62%, and 95.64%, respectively, with validation accuracies of 98.71%, 90.09%, and 92.95%. For the four aforementioned classifications, the MobileNetV3 model achieved accuracies of 95%, 100%, 100%, and 99%, respectively, and exhibited a precision of 98.25%, a recall of 98.22%, and an F1-score of 98.23%. Image analysis delineated boundaries between pathological regions and normal muscle tissues in WB, validated by bioimpedance and stress-strain measurements. Segmentation ranges for MILD, MOD, and SEV pathological severity were determined as approximately 55%, 62%, and 65%, respectively, marking the first precise internal stratification of WB. Results showed that ultrasound imaging combined with DL effectively assessed myopathy distribution within WB, enabling accurate classification and stratification for practical applications.