鎖骨骨折在所有成人骨折中極為常見，在一些緊急的情況下，如果沒有及時診斷和治療，將可能導致氣胸或出血等其他嚴重的併發症。所以能夠在情況較危急的急診室全天候正確有效的識別chest x-ray中危及生命的病例非常重要。然而，較為偏鄉的醫院通常沒有足夠的放射科醫生來全天候的工作以滿足這種即時的需求。另一方面，隨著人工智能(AI)的進步以及高端GPU(AI的關鍵創新)的成本下降，深度學習卷積神經網路的性能變得更加準確，並且目前常應用於各種汽車商業的用途。在這裡，我們假設這種基於GPU的新型AI技術可以轉移到醫學影像分析領域，為偏鄉醫院提供高精度的急性鎖骨骨折自動診斷服務。這項研究的數據集使用了4953張胸腔x-ray影像，其中左側及右側的鎖骨總共有9906根，這些數據集皆經由放射學專家評估和標記各個邊界框和骨折的種類(5287根正常的鎖骨、1068根急性的骨折、1656根舊傷的骨折和 1895根由骨釘固定的鎖骨)。全部的鎖骨圖像被分成6934(70%)張用於訓練 Mask R-CNN 模型和 2972張(30%)用於測試。最後測試的結果，急性病例檢測的召回率為0.781(所有急性病例共有324個案例，而預測為急性病例有253個案例)，檢測的精確度為0.821(預測為急性病例有308個，而有253個案例為急性病例)，最後F1的分數為0.801。;Clavicle fractures are extremely common in adult fractures and some emergency cases may cause other serious complications such as pneumothorax or hemorrhage if not diagnosed and treated timely. It is very important to be able to identify the life threatening cases in chest x-ray films correctly and efficiently around the clock in the emergency department. However, the rural hospitals generally do not have enough staff radiologists to work around the clock to meet this timely demand. On the other hand, with the advances in artificial intelligence (AI) and the cost drop of high end GPU (the key AI innovation), the performance of deep learning convolutional neural networks has become accurate enough and readily available for commercial use in various automobile applications. Here, we hypothesize that this novel GPU-based AI technology can be transferred to the field of medical image analysis for providing high accuracy services in rural hospitals for diagnosing acute clavicle fractures automatically. This study included a dataset of 4953 chest x-ray images with 9906 (left and right) clavicles evaluated and annotated by radiology specialists with bounding boxes and fractural types (5287 normal clavicles, 1068 acute fractures, 1656 chronic fractures, and 1895 fixations). The clavicle images were split into 6934 (70%) for training a Mask R-CNN model and 2972 (30%) for testing. The recall, precision, and F1-score of detection of acute cases in the test results were 0.781 (253 true positive acute calls over a total of 324 true positive acute cases), 0.821 (253 true positive acute calls over a total of 308 positive acute cases), and 0.801 respectively.
