利用U-Net卷積網路方法對電腦斷層影像進行肝腫瘤分割之研究

The Study of Segmented Liver Tumors from Computed Tomography by Using U-Net Convolutional Network Approach

李狄鴻(Ti-Hung Lee)

指導教授 ： 陳泰賓

共同指導教授 ： 杜維昌(Wei-Chang Du)

義守大學/電機資訊學院/資訊工程學系/碩士(2019年)

摘要

臨床上常利用電腦斷層影像進行肝臟切割、肝臟腫瘤體積計算、肝內血管瘤定量分析等。透過影像分割技術將肝臟器官或病灶進行區分可以應用於腫瘤分期及治療；然而該分割技術大多數採用手動或半自動方式進行分析，使得分析過程需花費較大時間成本。本研究利用深度學習網路架構，對電腦斷層肝臟腫瘤影像進行訓練，以建立自動肝臟腫瘤分割模型。
醫學影像由國際生物醫學影像研討會(International Symposium on Biomedical Imaging, ISBI)提供肝臟腫瘤影像；共有131位病患肝臟腫瘤3D腹部電腦斷層影像及標記影像，影像大小為512×512像素、影像格式為NIFTI；對每位病患3D腹部電腦斷層影像提取出2D肝臟腫瘤切面，共6792張2D肝臟腫瘤影像及其標記影像；標記影像具有肝臟及腫瘤二個部份以作為U-Net切割結果之金標準。每一張2D影像進行3種雙閾值處理後以24位元格式儲存成512×512×3之PNG格式；執行10次隨機選取訓練集與測試集進行U-Net建模同時輸入影像將調整成320×320×3之PNG格式；其中訓練集與測試集的影像張數比例分別為6792張之90%與10%。評估方法為測試集肝臟及腫瘤切割結果之Dice Score及準確度。
經由U-Net網路架構訓練之影像分割模型，於測試集肝臟及腫瘤Dice Score及準確度分別為96.04%及73.70%與96.22%及72.16%。
3種雙閾值處理轉成24位元影像可以提升網路架構學習Dice Score及準確度。未來可以考慮應用於臨床真實影像以及開發全三維影像切割之類神經網路模型。

關鍵字

影像分割 ； 肝臟腫瘤

並列摘要

The computed tomography (CT) is often used for segmentation, estimation of tumor volume, and quantitative analysis of hemangioma in clinics. Tumor staging and treatment were related image segmentation technology. However, it was time cost for manually or semi-automatically segmenting methods. In this study, a deep learning network approach was applied to automatically segment hepatic tumor in CT images.
CT hepatic images were obtained from the website of International Symposium on Biomedical Imaging (ISBI). Total of 131 patients were collected with 3D abdominal CT images including tumor labeled images. The size of slice was 512 × 512 pixels with NIFTI format. All slices including tumor were selected out from 3D abdominal CT image. Finally, there were 6792 2D slices including liver tumor and their labeled images. The golden of truth was those labeled images which were used to compare with segmented results by U-Net convolutional network approach. Three different dual thresholding were utilized to generate a 512×512×3 PNG format for each selected out 2D slice. Then, U-Net model was applied for arbitral selected 2D slices into training and test sets with ratio 9:1 and resizing image from 512×512×3 to 320×320×3. The evaluation of performance by U-Net was using Dice Score and accuracy in the test set.
The Dice Score of segmented between liver and tumor by using U-Net in the test set were 96.04% and 73.70%. The accuracy of segmented between liver and tumor by using U-Net in the test set were 96.22% and 72.16%.
The Dice Score and accuracy were improved by using three different dual thresholding methods with U-Net. Meanwhile, the presented method would be applied to real clinical images and developed fully 3D convolutional network approach.

並列關鍵字

Image Segmentation ； liver tumors ； U-Net

參考文獻

[1] Pham DL, Xu C, and Prince JL. Current methods in medical image segmentation. Annu Rev Biomed Eng. 2000; 2:315-37.

[2] Häme Y, and Pollari M. Semi-automatic liver tumor segmentation with hidden Markov measure field model and non-parametric distribution estimation. Med Image Anal.2012;16(1):140-9.

[3] Wu W, Wu S, Zhou Z, Zhang R, and Zhang Y. 3D Liver Tumor Segmentation in CT Images Using Improved Fuzzy C-Means and Graph Cuts. Biomed Res Int. 2017; 2017:5207685.

[4] Anter AM and Hassenian AE. CT liver tumor segmentation hybrid approach using neutrosophic sets, fast fuzzy c-means and adaptive watershed algorithm. Artif Intell Med. 2018; S0933-3657(18)30174-X.

[5] Long J, Shelhamer E, and Darrell T, Fully convolutional networks for semantic segmentation. Proceedings of the IEEE conference on computer vision and pattern recognition.2015; 3431-3440.