探討人工智慧方法於胸腔X光影像分類之表現

The Performance of Classified Chest X-Ray Images by Artificial Intelligence Methods

梁志宇(Chih-Yu Liang)

指導教授 ： 陳泰賓

共同指導教授 ： 杜維昌(Wei-Chang Du)

義守大學/智慧科技學院/資訊工程學系/博士(2022年)

摘要

動機與目的：
胸腔X光檢查在臨床上是重要及最普遍的檢查之一。胸部X光影像的診斷與判讀，取決於操作者暨醫師診斷之技術與經驗。近幾年人工智慧輔助影像類別判讀大多為針對標準造影影像；然而混合標準造影與住院病患影像之人工智慧深度學習判讀之研究仍不多見。因此本研究將探討混合不同造影標準之人工智能分類效能之差異。

材料與方法：
本研究使用之胸部X光影像取自美國國家衛生院(National Institutes of Health, NIH)公開資料庫胸部X光影像，從中篩選出五類常見胸部X光疾病；分別是積水(Effusion)、肺炎(Pneumonia)、浸潤(Infiltrate)、腫塊(Mass)及結節(Nodule)，各選72、120、123、84、79張總共478張。透過八種深度學習網路模型(Convolution Neural Network, CNN)進行影像特徵萃取，再由三種機器學習分別為支持向量機(Support Vector Machine, SVM)、邏輯斯特回歸(Logistic Regression)及單純貝氏分類器(Nave Bayes)建立分類模型；同時針對八種CNN採用遷移學習方式建立分類模型。驗證模型效能評估以測試組的準確性、特異性、靈敏度及 Kappa 一致性統計量。

結果：
局部和擴散之間的類別可以通過深度和機器學習方法和遷移學習成功分類，其準確度分別為 0.938 和 0.736；腫塊和結節組之間的最佳分類準確度為 0.912 和 0.912。 積水、浸潤組和肺炎組的最佳分類準確度為0.742 和 0.556 的準確度。 然而，分別使用兩種方法分類五個類別的準確度為0.458和0.417。

結論:
在這項研究中分類胸部 X 射線影像的兩種人工智能 (AI) 模式； 深度結合機器學習和遷移學習方法進行了比較，以比較五類、三類和兩類之效能；其中局部與擴散組透過深度和機器學習方法能成功分類。

關鍵字

人工智慧 ； 深度學習 ； 卷積神經網路 ； 影像分類 ； 胸部X光影像

並列摘要

Motivation and purpose:
Chest X-ray examination is one of the most important and common clinical examinations. The diagnosis and interpretation of chest X-ray images depend on the imaging technology and experience of interpreters. In recent years, the artificial intelligent methods were to assist classified categories of X-ray images with standard imaging protocol. However, the classified performance of artificial intelligent methods for hybrid standard imaging protocol and inpatient images is still rarely. Therefore, the performance of classification by using artificial intelligence for hybrid images was performed and investigated in this study.

Materials and methods:
The chest X-ray images used in this study were sampled from the National Institutes of Health (National Institutes of Health, NIH) chest X-ray images of public database. Five common chest diseases were including effusion, pneumonia, infiltrate, mass, and nodule with respectively to 72, 120, 123, 84 and 79 samples. A total of 478 chest X-ray images were involved in this study. The classified models were combined eight deep learning network models (Convolution Neural Network, CNN) for extracted features of images, and then three machine learning approaches including support vector machine (SVM), logistic regression (LR), and Nave Bayes (NB) were used to establish the classified model. Meanwhile, the transfer learning schema was applied to eight CNNs to train the classified models. The classified performance was evaluated by accuracy (ACC), specificity (SPE), sensitivity (Sensitivity, SEN), positive predictive value (PPV), negative predictive value (Negative Predicted Value, NPV) and Kappa agreement statistics on testing set.

Results:
The category between local and diffuse could be successfully classified by deep and machine learning approach and transfer learning with accuracy is 0.938 and 0.736. The best classified performance between mass and nodule groups with accuracy 0.912 and 0.912. The best classified performance among effusion, infiltrate, and pneumonia groups only with accuracy 0.742 and 0.556. However, the accuracy is 0.458 and 0.417 for classified five categories with respectively to two methods.

Conclusions:
In this study, the two kinds of artificial intelligent (AI) schemas are investigated for classified Chest X-ray images. The deep with machine learning and transfer learning methods had been compared with five, three, and two categories. The category between local and diffuse could be successfully classified by deep and machine learning approach.

並列關鍵字

artificial intelligence ； deep learning algorithms ； convolutional neural networks ； image classification ； chest X-ray images

參考文獻

[1] Baltruschat IM, Nickisch H, Grass M, Knopp T, Saalbach A. Comparison of Deep Learning Approaches for Multi-Label Chest X-Ray Classification. Sci Rep. 2019;9(1):6381.

[2] Liang G, Zheng L. A transfer learning method with deep residual network for pediatric pneumonia diagnosis. Comput Methods Programs Biomed. 2020;187:104964.

[3] Kumar K. Objective Image Quality Metrics for Ultrasound Imaging, Proceedings of ISSN: 2012; 2278-0181.

[4] Fukushima, K. Fukushima K. Neocognitron: a self organizing neural network model for a mechanism of pattern recognition unaffected by shift in position. Biol Cybern. 1980;36(4):193-202..

[5] Y. Lecun, L. Bottou, Y. Bengio and P. Haffner, Gradient-based learning applied to document recognition, in Proceedings of the IEEE, vol. 86, no. 11, 1998;2278-2324.