Berggreen Johan, Johansson Anders, Jahr John, Möller Sebastian, Jansson Tomas
Biomedical Engineering, Department of Clinical Sciences Lund, Lund University, Lasarettsgatan 37, 22185 Lund, Sweden.
Intensive and Perioperative Care, Skåne University Hospital, Entregatan 7, 22185 Lund, Sweden.
Healthcare (Basel). 2023 Jan 7;11(2):184. doi: 10.3390/healthcare11020184.
The number of hip fractures per year worldwide is estimated to reach 6 million by the year 2050. Despite the many advantages of regional blockades when managing pain from such a fracture, these are used to a lesser extent than general analgesia. One reason is that the opportunities for training and obtaining clinical experience in applying nerve blocks can be a challenge in many clinical settings. Ultrasound image guidance based on artificial intelligence may be one way to increase nerve block success rate. We propose an approach using a deep learning semantic segmentation model with U-net architecture to identify the femoral nerve in ultrasound images. The dataset consisted of 1410 ultrasound images that were collected from 48 patients. The images were manually annotated by a clinical professional and a segmentation model was trained. After training the model for 350 epochs, the results were validated with a 10-fold cross-validation. This showed a mean Intersection over Union of 74%, with an interquartile range of 0.66-0.81.
据估计，到2050年，全球每年髋部骨折的数量将达到600万。尽管区域阻滞在处理此类骨折疼痛方面有诸多优势，但与全身镇痛相比，其使用程度较低。一个原因是，在许多临床环境中，获得应用神经阻滞的培训和临床经验的机会可能是一项挑战。基于人工智能的超声图像引导可能是提高神经阻滞成功率的一种方法。我们提出一种使用具有U-net架构的深度学习语义分割模型来识别超声图像中股神经的方法。该数据集由从48名患者收集的1410张超声图像组成。这些图像由临床专业人员进行手动标注，并训练了一个分割模型。在对模型进行350个轮次的训练后，结果通过10折交叉验证进行验证。这显示平均交并比为74%，四分位间距为0.66 - 0.81。
