人工智慧輔助跟骨骨折X光判讀

Artificial Intelligence Assisted Calcaneal Fracture Interpretation on Radiographs

蕭名謙(Ming-Chian Hsiao)；李正輝(Cheng-Hui Lee)；侯嘉萍(Jia-Ping Hou)；呂坤木(Kun-Mu Lu)；楊必立(Pei-Li Yang)

人工智慧 ； 深度學習 ； X光 ； 跟骨骨折 ； Artificial intelligence ； Deep learning ； Radiograph ； Calcaneal fracture

中華放射線技術學雜誌

48卷1期（2024 / 03 / 01）

1 - 5

繁體中文;英文

跟骨骨折是跗骨骨折中最常見的類型，佔約60%；一般X光影像因可近性高、輻射劑量相對電腦斷層低且成本較低，是為跟骨骨折的醫學影像首要選擇。人工智慧近年在醫學影像有顯著進步。本研究目的為以深度學習卷積神經網絡訓練一人工智慧模型，用以自動化分類一般X光足踝側位照是否具有跟骨骨折。回溯性收入1126例一般X光足踝影像側位照，包含骨折299例和非骨折827例，經兩位放射診斷科醫師標註及影像前處理之後進行模型訓練。訓練、驗證、測試資料集以8:1:1分配。模型架構為ShuffleNet-v2-x0.5，並使用美國國立衛生研究院胸部X光資料集作為預訓練。結果在驗證資料集準確率達100.000%，測試資料集準確率達97.321%，真陽性為27例、真陰性為82例、偽陽性為0例、偽陰性為3例，曲線下面積為0.996，精確性為1.000、召回率0.900、F1分數為0.947。證實深度學習的卷積網絡得以在醫學影像有優良表現，該模型可能作為初步篩檢或輔助診斷以協助放射線科醫師的影像判讀工作。

Calcaneal fracture is the most common type of tarsal fractures, accounting for about 60%; calcaneus radiograph is the primary medical imaging choose for calcaneal fractures due to high accessibility, lower radiation dose compared to computed tomography, and lower cost. Artificial intelligence has made significant progress in medical imaging in recent years. The purpose of this study is to train an artificial intelligence model with a deep learning convolutional neural network to automatically classify whether there is a calcaneal fracture in lateral view of ankle radiograph. Retrospective collection of 1126 cases of lateral view of ankle radiographs, including 299 cases of fractures and 827 cases of non-fractures, was annotated by two radiologists and imaging pre-processed for model training. The training, validation, and test datasets are allocated 8:1:1. The model architecture is ShuffleNet-v2-x0.5, and the chest X-ray data set of the National Institutes of Health is used as pre-training. The accuracy rate in the validation data set is 100.000%, and the accuracy rate in the test data set is 97.321%. There are 27 true positives, 82 true negatives, 0 false positives, and 3 false negatives. The area under the curve was 0.996. The precision is 1.000, the recall is 0.900, and the F1 score is 0.947. It is confirmed that the convolutional network of deep learning can perform well in medical imaging. This model may be used as a preliminary screening or auxiliary diagnosis to assist radiologists in image interpretation.