利用深度學習之胸部X光影像自動化肺臟分割

Automatic Lung Segmentation in Chest X-ray Images using Deep Learning

林依嫺(Yi-Hsien Lin) ； 湯韻璇(Yun-Xuan Tang) ； 葉士詠(Shih-Yung Yeh) ； 彭惠玲(Hui-Ling Peng) ； 陳良光(Liang-Kuang Chen) ； 蔡裕豐(Yuh-Feng Tsai) ； 施政廷(Cheng-Ting Shih) ； 楊必立(Pei-Li Yang)

《臺灣醫事放射期刊》 11卷1期 (2023/07) Pp. 10-16

https://doi.org/10.6717/JTMRT.202307_11(1).0002

摘要

現今已有大量研究發展深度學習工具於胸部X光影像上診斷肺臟疾病，並展現出極佳的診斷準確率及效率，然而，其診斷效能將受到肺臟以外區域的解剖構造或其他物件影響，因此，本研究利用深度學習建構胸部X光影像自動化肺臟分割工具。本研究基於ResUnet建構胸部X光肺臟分割卷積神經網路（lung segmentation CNN, LS-CNN），並將損失函數改為正規化戴斯損失，以平衡肺臟與背景區域大小的不同，提升分割結果的準確性，本研究使用從公開資料庫收集6000例胸部X光影像及肺臟標準遮罩，分別以4800例及1200例進行LS-CNN的訓練及驗證，並計算分割準確率、平均重疊度（intersection over union, IoU）及輪廓匹配分數（contour matching score, BF score）量化評估分割效能。在驗證案例中，LS-CNN之分割準確率、IoU及BF score平均皆高於0.98，分割效能亦不受姿勢不同、非解剖物件、患者體型、年齡影響，顯示LS-CNN之強健性。本研究提出之LS-CNN可望應用於協助胸部X光診斷工具之開發及臨床診斷，以提升胸部X光臨床診斷之品質及效率。

關鍵字

胸部X光 ； 深度學習 ； 肺臟分割

並列摘要

A large number of studies have developed deep learning tools to diagnose lung diseases on chest X-ray images, and have shown excellent diagnostic accuracy and efficiency. However, the diagnostic performance of the tools is strongly influenced by the anatomical structure or other objects outside the lungs. Therefore, this study developed a deep learning-based tool for automatic lung segmentation from chest X-ray images. In this study, a chest X-ray lung segmentation convolutional neural network (LS-CNN) was constructed based on the ResUnet. In addition, a normalized Dice loss was applied as the loss function for the training of the LS-CNN to improve the segmentation accuracy by balancing the difference between the area size of the lung and the background. 6000 cases of chest X-ray images and lung standard masks were collected from public databases. 4800 cases and 1200 cases were used to train and validate the LS-CNN, respectively. The segmentation performance was quantitatively evaluated by calculating the accuracy, intersection over union (IoU), and contour matching score (BF score) of the segmentation results. In the validation set, the segmentation accuracy, IoU, and BF score of LS-CNN are all higher than 0.98 on average. The segmentation performance of the LS-CNN is not affected by different postures, non-anatomical objects, patient size, and age, which indicates the robustness of LS-CNN. In conclusion, the proposed LS-CNN could be applied to assist the development of chest X-ray diagnostic tools and clinical diagnosis, to improve the quality and efficiency of the clinical diagnosis of chest X-ray images.

並列關鍵字

chest X-ray ； deep learning ； lung segmentation