運用深度學習做肺結核之特徵分類器

研究背景與目的結核病(Tuberculosis)是目前仍然普遍存在於全世界，尤其是在未開發及開發中國家的慢性傳染病，每年都有許多人因該病而死亡，依目前現行的醫療技術，若及早發現並給予適當的抗結核藥物治療，是可以幾乎百分之百的控制和痊癒的，但其臨床病徵不明顯，導致診斷結核病仍然是一項挑戰，而有關影像辨識的深度學習模型正在快速的發展當中，但並非每一個影像辨識的模型適合結核病，需要進一步試驗探索在不同模型的特性與合適與否。本篇研究將利用現有的深度學習結構，製作一個有關肺結核的特徵分類器。研究方法建立並訓練三個獨立且不同結構之深度學習模型分別為GoogLenet、VGG16、VGG19，並使用遷移學習及兩個肺結核胸部Ｘ光片的開放性資料集MontgomerySet和ChinaSet訓練(共800張)。將模型串連後成特徵分類器，並使用其提取之特徵進行圖片預測，進行驗證與比較。研究結果三個模型VGG19、VGG16和GoogLeNet的AUC值分別為0.9、0.87、0.52，表示VGG模型都有良好的鑑別力;而GoogLeNet的ROC呈現剛好為對角線，表示模型無鑑別力。製作特徵分類器的部分則將VGG16與VGG19做串連，而在與單一模型進行預測分析比較時，預測圖片的正確率表現為VGG19最優（87%），特徵分類器次之（85%），VGG16最後（83%）。結論與建議本研究提供一個利用兩個已訓練的VGG的串連模型可以進行快速的圖像預測分析，未來可以通過特徵分類器有效整合已訓練之模型，進行多模型快速預測分析且多重驗證;而在學術研究上，則能利用特徵分類器探討分析數據的影響因子與深度模型之間的相關統計學研究。

關鍵字

深度學習；肺結核；影像辨識；分類器

並列摘要

Background: Tuberculosis is a chronic infectious disease that is still prevalent throughout the world, especially in undeveloped and developed countries. Many people die every year because of the disease. If tuberculosis is discovered in early stage and given appropriate anti-tuberculosis drugs, it can be almost 100% controlled and cured with current medical technology. However, their clinical signs are not obvious, leading to the diagnosis of tuberculosis remains a challenge. The deep learning model for image recognition is rapidly developing, but not any image-recognition model is suitable for tuberculosis. Further experiments are needed to explore the characteristics and suitability of different models. This study used the existing deep learning structure to create a feature classifier for tuberculosis. Methods: Three independent and different deep learning models, GoogLenet, VGG16, and VGG19, were established and trained. Two open datasets of tuberculosis chest X-rays, MontgomerySet and ChinaSet (800 sheets), were used to train the model. The model is concatenated into a feature classifier. The extracted features were used for image prediction, and verification and comparison. Results: The AUC values of the three models VGG19, VGG16 and GoogLeNet are 0.9, 0.87, and 0.52, respectively. It is obvious that the VGG model has good discriminative power. The GoogLeNet has no discriminative powered because of its ROC curve is just diagonal such the proposed feature classifier is connected in series with VGG19. Compared with the single model for predictive analysis, the correct rate of VGG19 is best (87%). The proposed feature classifier is second (85%). The VGG16 is worst (83%). Conclusion and suggestion: This study provides a fast model for predictive analysis using two trained VGG concatenation models. In the future, proposed feature models can be used to effectively integrate trained models, to perform multi-model fast prediction analysis and multiple verifications. In the research, the proposed feature classifier can be used to explore the correlation between the impact factor and the depth model of the analytical data.

並列關鍵字

deep learning ； tuberculosis ； image recognition ； classifier

參考文獻

Alom, M.Z., et al., The history began from AlexNet: a comprehensive survey on deep learning approaches. arXiv preprint arXiv:1803.01164, 2018.

Google Scholar

Bar Y, Diamant I, Wolf L, Greenspan H. Deep learning with non-medical training used for chest pathology identification. In: Hadjiiski LM, Tourassi GD, eds. Proceedings of SPIE: medical imaging 2015—computer-aided diagnosis. Vol 9414. Bellingham, Wash: International Society for Optics and Photonics, 2015; 94140V.

Google Scholar

Fukushima, K. (1988). Neocognitron: A hierarchical neural network capable of visual pattern recognition. Neural Networks, 1(2), 119–130. doi: 10.1016/0893-6080(88)90014-7

Google Scholar

Hubel, D., & Wiesel, T. (1968). Receptive fields and functional architecture of monkey striate cortex. J Physiol, 195(1), 215–243.

Google Scholar

He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. arXiv pre-print. https://arxiv.org/abs/1512.03385. Published December 10, 2015. Accessed Sep-tember 20, 2016

Google Scholar

國際替代計量

運用深度學習做肺結核之特徵分類器

未授權

主題瀏覽