Abstract: Upper limb fractures are common yet challenging traumatic injuries in clinical practice, where diagnostic accuracy is crucial for effective treatment and patient recovery. The traditional X-ray-based fracture diagnosis method is cumbersome, time-consuming, and difficult to meet the high demands of modern medical imaging in terms of efficiency and accuracy. Against this backdrop, deep learning-assisted diagnosis of upper limb fractures primarily leverages deep learning models for classification, detection, and segmentation of medical images, enabling the identification of abnormalities within the images. This approach enhances the speed and accuracy of diagnostic models while also providing clinicians with valuable auxiliary insights. To gain a comprehensive understanding of the current research status and advancements in deep learning techniques for upper limb fracture diagnosis, this paper first provides a detailed overview of several common types of upper limb fractures and summarizes widely used public datasets in this field. Simultaneously, it systematically reviews commonly employed evaluation metrics, facilitating a more nuanced understanding of model performance across various tasks. Secondly, an in-depth analysis is conducted on the application progress of deep learning in the three major computer vision tasks: image classification, object detection, and image segmentation. A detailed comparative analysis is performed on the primary optimization strategies of different algorithms, the issues they address, and their advantages? and limitations. Furthermore, a comprehensive summary of the interpretability of deep learning models is provided. Finally, this paper provides a comprehensive comparison in terms of dataset size, methodologies, algorithm advantages and disadvantages, and experimental results. It systematically summarizes the key challenges currently faced in upper limb fracture diagnosis and offers prospects for future research directions.

Key words: deep learning, upper limb fractures, convolutional neural networks, computer vision (CV)

摘要： 上肢骨折作为临床常见且复杂的创伤性骨折类型，其诊断准确性对患者的治疗和康复具有重要的意义。传统X射线（X-ray）诊断方法操作繁琐且耗时，难以满足现代医学对高效、精确诊断的需求。在此背景下，深度学习辅助上肢骨折诊断主要利用深度学习模型对医学影像进行分类、检测和分割，确定影像中是否存在异常，提高模型诊断的速度和准确性，同时也为医生提供了更有价值的辅助意见。为了更好地了解深度学习技术在上肢骨折诊断领域的研究现状和进展，详细介绍了几种常见的上肢骨折类型并对当前广泛使用公开的上肢骨折数据集进行了总结，对常用的评价指标进行了系统归纳，以便更好地理解模型在不同任务中的性能表现。深入分析了深度学习在图像分类、目标检测和图像分割三项计算机视觉（CV）任务中的应用进展，通过总结各算法的优化策略及其在骨折诊断中的具体应用，比较了各自的优势与局限性，并对深度学习模型的可解释性进行总结。从数据规模、使用方法、算法优缺点及实验结果等方面进行了全面对比，系统总结了当前上肢骨折诊断中面临的主要挑战，并对未来研究方向进行了展望。

关键词: 深度学习, 上肢骨折, 卷积神经网络, 计算机视觉（CV）