應用遮罩區域型卷積神經網路之技術於電腦斷層掃描影像之肝腫瘤細胞偵測與分割最佳化

An Optimization-based Technique Applied by Mask RCNN for Liver Tumor Detection and Segmentation on CTScan Images

林均泰(Chun-Tai Lin)

指導教授 ： 林永松

國立臺灣大學/管理學院/資訊管理學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201902824

摘要

肝癌已成為全球癌症死因中的主要原因，甚至是台灣地區男性癌症的第一名主要死因以及女性癌症的第二名主要死因。為了利於提早進行肝癌的治療，癌症診斷的準確率以及效率就顯得非常重要。傳統的診斷方法需要耗時費力地透過人工來辨識影像，不僅降低了醫生們的工作效率，也連帶地影響了醫療品質。因此，本研究 擬提出了一套方法以自動化的對電腦斷層掃描影像 (CT) 進行影像辨識，並標記出肝癌細胞的區域，以輔助醫生進行肝癌的診斷。本論文 訓練了兩個 Mask R-CNN的模型，分別用來辨識出肝臟以及肝癌細胞的位置，並把這兩個模型結合在一起。第一個 Mask R-CNN的模 型會將傳入的電腦斷層掃描影像中所包含肝臟的區域 偵測並標記切割出來，作為RoI傳入到第二個模型；而第二個 Mask R-CNN模型則會將前一個模型所標記出的 RoI區域內包含肝癌細胞的區域 同時偵測並標記切割出來。在訓練 Mask R-CNN 的模型時，我們特別加重圖像切割的損失函數的權重，並且也應用懲罰矩陣來加重被分類錯誤的像素權重。本研究的訓練以及測試資料是使用 MICCAI 2017 LiTS所提供的公開資料，其中包含130組病人的腹腔電腦斷層掃描影像 並且 在肝與肝腫瘤的圖像切割的 Dice 分別達到 94.1%和75.2%。

關鍵字

區域型卷積神經網路 ； 肝臟 ； 肝癌細胞 ； 物體偵測 ； 影像分類 ； 圖像切割

並列摘要

Liver cancer has become one of the leading cause of death worldwide. While focusing on Taiwan, liver cancer is the leading cause of death from cancer in males and the second leading cause in females. In order to determine the treatment options earlier, the accuracy and efficiency play important roles in liver cancer diagnosing. The traditional way to diagnose liver cancer is to identify whether or not the CT slices contain the tumor manually, which will reduce the working efficiency for doctor and affect the healthcare quality. As a result, this study proposes a framework to automatically segment liver and its lesion in CT volumes for assisting doctors to efficiently diagnose liver cancer. We train and cascade two Mask R-CNN models which are used to segment liver and lesions separately. The first model of Mask R-CNN will segment liver from the CT slices as RoI input for the second model. The second Mask R-CNN will segment lesions with the predicted liver RoIs from previous model. While training the Mask R-CNN model, we enhance the weight of segmentation loss function and apply penalty metrics to add on weights on the pixels classified incorrectly. We train on a public dataset which contains 130 abdominal CT volumes of patients offered by MICCAI 2017 LiTS. The framework proposed by the paper achieves Dice score 94.1% and 75.2% respectively for liver and liver tumor segmentation.

並列關鍵字

R-CNN ； Liver ； Tumor ； Classification ； Detection ； Segmentation

參考文獻

[1] F. Bray, J. Ferlay, I. Soerjomataram, R. L. Siegel, L. A. Torre, and A. Jemal, "Global Cancer Statistics 2018: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries," A Cancer Journal for Clinicians, vol. 68, no. 6, pp. 394-424, November 2018.

[2] Department of Health (DOH): Vital statistics, “Taiwan Public Health Report”, the Executive Yuen, May 2017.

[3] X. Huang, and G. Tsechpenakis, “Medical Image Segmentation,” Information Discovery on Electronic Health Records, pp. 251-289, June 2009.

[4] J. Stawiaski, E. Decenciere, and F. Bidault, “Interactive Liver Tumor Segmentation Using Graph-cuts and Watershed,” Medical Image Computing and Computer Assisted Intervention, New York City, USA, September 2008.

[5] D. Wong, J. Liu, Y. Fengshou, Q. Tian, W. Xiong, J. Zhou, Y. Qi, T. Han, S. Venkatesh, and S. C. Wang, “A Semi-Automated Method for Liver Tumor Segmentation Based on 2D Region Growing with Knowledge-based Constraints,” The MIDAS Journal, vol. 41, no. 43, pp. 159-162, July 2008.