Myofascial pain syndrome (MPS) is a Myofascial strain disease characterized by myofascial trigger points (MTrPs), presenting with local pain, dysfunction, and decreased muscle strength. Epidemiological data show that the proportions of visits to basic medical care and pain clinics are 15% and 90%, respectively. The incidence rate in China ranges from 30% to 93%, and the difference is due to the non-uniform diagnostic criteria. At present, clinical diagnosis mainly relies on doctors' palpation, which is highly subjective and lacks objective detection methods. With the changing modern lifestyle (such as long-term head-down posture), MPS is showing a trend of younger onset, but public awareness and diagnostic systems remain insufficient. MTrPs, as a core pathological feature, are prone to occur in the neck, shoulders, and back. During its active period, a vicious cycle of "pain - limited movement- new trigger point" can be formed. Ultrasonic technology has become a promising diagnostic tool due to its advantages, such as real-time, non-invasive imaging, but is constrained by factors including operator experience and equipment performance. In recent years, deep learning, especially transfer learning and convolutional neural network models, has demonstrated outstanding performance in medical image analysis, capable of automatically extracting features from ultrasound images and providing objective decision support.

To explore the application effect of a musculoskeletal ultrasound image analysis based on a deep convolutional neural network (DCNN) in the intelligent diagnosis of MPS.

Five hundred patients with neck, shoulder, and back pain who visited our hospital from May 2021 to June 2025 were selected as the research subjects, and their musculoskeletal ultrasound image data were collected. After eliminating unqualified images, 1,692 ultrasound images from 481 patients were included to construct an image dataset, and three attending physicians labeled them. The dataset was split into a training set and a test set, each containing 1 383 and 309 musculoskeletal ultrasound images, respectively. Deep learning algorithms were used to construct diagnostic models (DCNN1, DCNN2, DCNN3) to assess their accuracy in identifying active and latent MTrPs. In addition, 200 musculoskeletal ultrasound images of patients with neck, shoulder, and back pain, independent of the database, were selected and evaluated by three chief ultrasound physicians. The recording took a long time and was compared with the model diagnosis results. Another 130 independent musculoskeletal ultrasound examination images of patients with neck, shoulder, and back pain were selected from the database. Six ultrasound resident physicians evaluated the images before and after model assistance, respectively, and the diagnostic results were compared.

The accuracy of the DCNN3 model in diagnosing active and latent MTrPs was 0.983 and 0.876, respectively. The model's accuracy in identifying active and latent MTrPs was comparable to the diagnostic level of three chief physicians, and the time the model took to evaluate images was significantly shorter than that of the chief physicians (P < 0.05). With the model's assistance, the accuracy of image diagnosis by resident physicians was significantly improved, and the time required to diagnose images was significantly reduced (P < 0.013) .

This intelligent assessment model has high accuracy in diagnosing active and latent MTrPs and can significantly improve diagnostic efficiency. It can effectively assist clinicians in correctly evaluating MPS in the neck, shoulders, and back.