應用深度學習於肺結節CT影像醫療診斷報告自動化生成技術

Application of Deep Learning for Medical Image Report Generation Technology in Lung Nodule CT Images

洪聖傑(Sheng-Chieh Hung)

指導教授 ： 曾明性

中山醫學大學/健康管理學院/醫學資訊學系/碩士(2021年)

https://doi.org/10.6834/csmu202100072

摘要

肺癌長年位居台灣癌症死亡原因之首，108年氣管、支氣管和肺癌為我國國人十大癌症死亡率的第一名，抽菸以及空氣汙染等問題同時也在增加罹患肺癌的機率，若能夠早期進行肺癌檢測並盡速接受治療，便能幫助降低肺癌病患的死亡率。
肺癌主要分為小細胞癌與非小細胞癌，非小細胞癌約佔全部肺癌百分之八十，若使用低劑量胸部電腦斷層便能夠在非侵入的情況下快速檢查，在當今深度學習竄起的時代，應用電腦科技去辨識病患影像成為一種新型態的方式，並且數據分析競賽平台提供了大量的病患肺部CT影像資料，讓全球能夠使用大量的公開影像集進行深度學習的訓練與應用。
在本研究中，使用HSMCNN三維網路模型對肺部結節影像集進行訓練，透過訓練裁切的結節影像將結節影像處理為特徵資料型態，並使用多類別語義標籤對結節特徵進行診斷文本的生成訓練，接著以編碼器-解碼器框架將經過訓練之三維網路模型作為編碼器，探討使用不同之遞迴神經網路作為解碼器進行醫療診斷字幕生成，可協助放射科醫生做快速的輔助診斷，而生成之醫療診斷語句將使用BLEU、ROUGE、METEOR、CIDEr四項評估指標進行效能評估，最終指標測試可達95%以上。

關鍵字

肺癌 ； 深度學習 ； 字幕訓練 ； CT影像 ； 神經網路

並列摘要

Lung cancer is a common cause of cancer deaths. In Taiwan, trachea, bronchus and lung cancer are the top ten cancer deaths in 2019. If the signs of lung cancer can be found in the early stage through research and treatment, the mortality rate of patients can be greatly reduced. Lung cancer is mainly divided into small cell carcinoma and non-small cell carcinoma. And the non-small cell carcinoma accounts for about 80% of all lung cancers. If low-dose chest computed tomography is used, it can be quickly examined in a non-invasive manner. In today's era of deep learning, the application of computer technology to identify patient images has become a new type of method. The data analysis competition platform provides a large amount of patient lung CT image data, allowing the world to use a large number of public image collections. Conduct deep learning training and application. In this study, the HSMCNN three-dimensional network model is proposed to train the lung nodules image set. The nodule image is processed into a feature data type through training and cropped nodule images, and multi-category semantic tags are used to generate diagnostic text for the nodule features. Then use the encoder-decoder framework to use the trained 3D network model as the encoder, and explore the use of different recurrent neural networks as the decoder for medical diagnosis subtitle generation. Try to assist the radiologist make a quick auxiliary diagnosis, and the generated medical diagnosis sentence will use four evaluation indicators, BLEU, ROUGE, METEOR, and CIDEr, for performance evaluation, and the final indicator test can reach more than 95%.

並列關鍵字

Lung cancer ； Deep learning ； Subtitle training ； Computed tomography ； Neural network

參考文獻

[1] Anil, G., Hegde, A., Chong, F. V. (2011). Thyroid nodules: risk stratification for malignancy with ultrasound and guided biopsy. Cancer Imaging, 11(1), 209.

[2] Armato III, S. G., McLennan, G., McNitt-Gray, M. F., Meyer, C. R., Yankelevitz, D., Aberle, D. R., . . . MacMahon, H. (2004). Lung image database consortium: developing a resource for the medical imaging research community. Radiology, 232(3), 739-748.

[3] Armato, S. G., 3rd, McLennan, G., Bidaut, L., McNitt-Gray, M. F., Meyer, C. R., Reeves, A. P., . . . Croft, B. Y. (2011). The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): a completed reference database of lung nodules on CT scans. Med Phys, 38(2), 915-931. doi:10.1118/1.3528204

[4] Banerjee, S., Lavie, A. (2005). METEOR: An automatic metric for MT evaluation with improved correlation with human judgments. Paper presented at the Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization.

[5] Camarlinghi, N., Gori, I., Retico, A., Bellotti, R., Bosco, P., Cerello, P., . . . Peccarisi, M. (2012). Combination of computer-aided detection algorithms for automatic lung nodule identification. International journal of computer assisted radiology and surgery, 7(3), 455-464.