電腦輔助診斷X光片胸腔疾病透過具圖注意力機制的卷積神經網路

Computer-Aided Diagnosis of X-ray Thorax Diseases Using CNN Model with Graph Attention Mechanism

黃聖凱(Sheng-Kai Huang)

指導教授 ： 張瑞峰

國立臺灣大學/電機資訊學院/資訊網路與多媒體研究所/碩士(2020年)

https://doi.org/10.6342/NTU202003136

摘要

胸部X光在臨床上是常用於診斷胸腔疾病的診斷工具之一，而胸腔疾病是造成世界各地生命損失的常見原因。為了降低胸腔疾病造成的死亡率，及早的發現與治療是非常重要的。電腦輔助診斷(Computer-aided Diagnosis, CAD)系統可以幫助放射科醫生識別胸腔疾病，進而提高放射科醫生的診斷準確度，並有助於及早開始治療過程。因此，我們提出一個電腦輔助診斷系統使用胸腔X光片來診斷胸腔疾病，透過卷積神經網絡(Convolutional Neural Network, CNN)擷取的影像特徵以及利用含圖注意力機制(Graph Attention Mechanism)的圖神經網路(Graph Neural Network, GNN)強化不同疾病之間的關係來改善診斷表現。在實驗中，我們利用來自於公開資料集(NIH Chest X-ray dataset)中共112,120張正面胸腔X光片，在此資料集中包含14種常見胸腔疾病標籤，而在這14種常見的胸腔疾病中每一張X光片可具有多個不同標籤，來評估提出的方法的表現。根據實驗結果，我們提出的方法表現優於先前研究相關的方法，在14種常見的胸腔疾病的平均曲線下區域(AUC)分數達到0.8266，而平均的準確度、靈敏度和特異性分別為0.7504，0.7704和0.7495。我們提出的方法整合了卷積神經網絡模型和包含圖注意力機制的圖神經網路模型並利用到疾病之間加強後的關聯性來改善診斷表現，而提出的方法可以有效的診斷胸腔疾病並有較好的診斷表現。

關鍵字

胸部X光 ； 胸腔疾病 ； 電腦輔助診斷 ； 圖神經網路 ； 圖注意力機制 ； 卷積神經網路

並列摘要

The chest x-ray is one of the most common techniques in clinical for thorax disease diagnosis, and thorax diseases are common causes of global life loss. For thorax patients, early diagnosis and treatment could reduce the mortality rate. A good computer-aided diagnosis (CAD) system could not only help radiologists to recognize different thorax diseases but also improve the diagnosis performance. Hence, we proposed a thorax CAD system using chest x-ray images through the convolution neural network (CNN). The proposed CAD system extracted the image representation features of chest x-ray images and further employed the graph neural network (GNN) with graph attention mechanism to enhance the correlation between different diseases to improve thorax disease diagnosis performance. In the experiments, 112,120 frontal-view chest x-ray images with labels of 14 common thorax diseases from an open dataset (NIH Chest X-ray dataset) were used to evaluate the performance of the proposed method, each chest x-ray image could have multiple thorax diseases label. Based on the experimental results, the best diagnosis performance of our method is better than previous related works that the average AUC score of 14 common thorax diseases is 0.8266, and the average accuracy, sensitivity, and specificity are 0.7504, 0.7704, and 0.7495, respectively. In summary, the proposed method integrates the CNN model and the GNN model with the graph attention mechanism to diagnose thorax diseases efficiently and has better diagnosis performance.

並列關鍵字

Chest x-ray ； Thorax diseases ； Computer-aided diagnosis ； Graph neural network ； Graph attention mechanism ； Convolution neural network

參考文獻

[1] A. K. G. M. Bekas, K. Pachocki, “How often are X-rays used as diagnostic tool by healthcare providers in the Mazovian Province of Poland,” Rocz Panstw Zakl Hig, pp. 155-160, 2013.

[2] G. C. o. D. Collaborators, “Global, regional, and national age-sex-specific mortality for 282 causes of death in 195 countries and territories, 1980–2017: a systematic analysis for the Global Burden of Disease Study 2017,” The Lancet, vol. 392, pp. 1736-1788, 2018.

[3] N. G. Csikesz, and E. J. Gartman, “New developments in the assessment of COPD: early diagnosis is key,” International journal of chronic obstructive pulmonary disease, vol. 9, pp. 277-286, 2014.

[4] S. F. Nemec, A. A. Bankier, and R. L. Eisenberg, “Pulmonary Hyperlucency in Adults,” American Journal of Roentgenology, vol. 200, no. 2, pp. W101-W115, 2013/02/01, 2013.

[5] M.-A. Blanchette, and J.-M. Grenier, “Subtle radiographic presentation of a pleural effusion secondary to a cancer of unknown primary: a case study,” The Journal of the Canadian Chiropractic Association, vol. 58, no. 3, pp. 273-279, 2014.