System for rapid measurement of muscle fiber morphology based on the FastSAM model

Muscle fibers can represent the fundamental structural units of the skeletal muscle. Their morphology is closely linked to the functional performance, adaptability to mechanical stress, and quality traits in both biomedicine and agriculture. During resistance training, the skeletal muscle can experience mechanical tension, which can induce microscale damage, followed by repair and remodeling that enlarge the cross-sectional area (CSA) of the fibers. This adaptive phenomenon can often be referred to as supercompensation, thereby underlying the muscle hypertrophy, strength gains, and enhanced tolerance. In livestock science, the fiber number, CSA, and diameter are the most critical determinants of the meat quality, thus influencing the pH, tenderness, protein composition, and water-holding capacity. Smaller fibers with the lower CSA are associated with improved tenderness, whereas the enlarged fibers often reduce the meat quality. In medical research, the quantitative analysis of the fiber morphology can provide the diagnostic markers for the neuromuscular disorders, muscle atrophy, and therapeutic efficacy, with the CSA reduction serving as a hallmark of the cachexia and age-related degeneration. hematoxylin-eosin (HE) staining, a cost-effective and reproducible approach, can be commonly employed to visualize the fiber boundaries and morphology, particularly for the histological evaluation of the muscle fibers. However, manual or semi-automatic quantification of the fiber size and density is labor-intensive and prone to subjectivity, particularly in the high-density tissue sections. Recent advances in computer vision and deep learning have introduced segmentation models, such as Cellpose, StarDist, RetinaMask, and FeatureNet. The efficiency can be improved to maintain the high accuracy for the complex muscle structures. segment anything model (SAM) has also attracted much attention for its zero-shot generalization. Yet, its computational demands have hindered the large-scale deployment. The inference speed and flexibility can also be offered using improved lightweight derivatives, such as the fast segment anything model (FastSAM), mobile segment anything model (MobileSAM), and edge segment anything model (EdgeSAM). In this study, an automated pipeline was developed to quantify the muscle fibers using FastSAM, which integrated a YOLOv8-segmentation backbone with the optimal pre- and post-processing. Pig longissimus dorsi samples were collected over the eight developmental stages, and then fixed, embedded, sectioned, and stained with HE. Images were acquired at 10× magnification using a Zeiss upright microscope. The fiber boundaries were enhanced after grayscale conversion, histogram equalization, and Laplacian sharpening. Large images were systematically cropped into the overlapping sub-regions to preserve the fiber integrity, followed by the batch inference on an NVIDIA L40 graphics processing unit. The regions obscured by background noise were removed to reconstruct the fiber boundaries. The CSA and Feret diameters (minimum, maximum, and average values) were also calculated after operation. The dataset was standardized in the COCO format to facilitate reproducibility and integration with the various platforms. An in-browser interface (which was containerized via Docker and supported by NGINX) was implemented to enable the user-friendly image upload, batch processing, visualization, and data export. A comparison was also made to evaluate the performance of the SAM, MobileSAM, and EdgeSAM. The results demonstrated that the superior performance of FastSAM was achieved with the higher Dice coefficient, precision, recall, and F1 score, while the lower inference was maintained, compared with SAM. Correlation analysis revealed that there was a strong agreement between automatic and manual measurements (r > 0.99), indicating the high reliability of the system. Twelve muscle groups were applied to highlight the high heterogeneity in the CSA and diameter distributions, indicating the functional specialization and fiber type composition. Developmental analysis over the weight stages confirmed that the progressive enlargement of the fibers was consistent with the growth physiology. In conclusion, A scalable platform was established for the biomedical diagnostics, livestock breeding, and meat quality assessment, bridging computation with the applied muscle biology. The FastSAM framework can provide a robust, efficient, and accurate solution to quantify the muscle fiber morphology in the HE-stained sections. Furthermore, the high-throughput analysis can also offer standardized outputs for the downstream research. The immunohistochemical markers and multimodal learning can be integrated to differentiate the fiber types in the future.