應用多階段U-Net優化模型於電腦斷層影像進行腎臟腫瘤分割

An Optimization-based Multi-Stage U-Net Model on Computer Tomography Images for Kidney Tumor Segmentation

鄭捷云(Chieh-Yun Cheng)

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(2022年)

https://doi.org/10.6342/NTU202203711

摘要

在腎臟腫瘤的醫療診斷上，醫生往往需要透過電腦斷層掃描進行判斷。傳統上，醫療影像的判讀高度依賴醫生的臨床經驗與專業知識，再加上電腦斷層掃描的大量切片與低對比度導致醫生需要花費大量的時間與精力。隨著影像辨識技術的進步，開始有許多自動分割的技術產生從而輔助醫生提升診斷的效率。
本文提出一種五階段架構設計，引入了EfficientNet with noisy student作為U-Net編碼器。在第一階段分割腎臟區域，並以其為基礎在第二階段分別訓練可以針對不同腫瘤特色的模型。在第三階段綜合第二階段中三個模型的優點進行加權投票從而提升腫瘤分割的表現。接著在第四階段檢查預測結果，降低偽陰性的發生。最後，在第五階段補充腫瘤的端點區域。
我們使用2019 Kidney Tumor Segmentation Challenge (KiTS19) 資料集進行實驗，該資料集包含210位有標記的病患與90位未標記的病患。在KiTS19官網中，我們得到腎臟分割的dice分數為0.9660，腫瘤分割的dice分數為0.7407。在本研究中，我們主要有兩個貢獻。第一點為提升2D模型在腎臟與腫瘤分割的表現。第二點為加快預測速度和降低所需的計算資源，本架構可以使模型靈活應用於資料有限或運算資源有限的場景中。另外，我們以本論文所提出的架構建立了系統，希望通過這個系統讓醫生在臨床診斷更為便利。

關鍵字

影像分割 ； 醫療影像 ； 腎臟 ； 腫瘤 ； 2D U-Net

並列摘要

In the medical diagnosis of kidneys and tumors, doctors often need to judge through computer tomography (CT). Traditionally, the interpretation of medical images is highly dependent on the clinical experience and professional knowledge of doctors. Coupled with a large number of slices and low contrast of the computed tomography scan, the doctor must spend a lot of time and energy. With the advancement of image segmentation technology, many automatic segmentation technologies have begun to be produced to assist doctors in improving the efficiency of diagnosis.
This thesis proposes a five-stage architecture design and introduces EfficientNet with noisy student as U-Net encoder. The kidney region is segmented in the first stage. Based on it, models that can target different tumor characteristics are trained separately in the second stage. In the third stage, the advantages of the three models in the second stage are combined for weighted voting to improve the performance of tumor segmentation. The predicted results are then checked in the fourth stage to reduce the occurrence of false negatives. Finally, the endpoint region of the tumor is replenished in the fifth stage.
We performed experiments using the 2019 Kidney Tumor Segmentation Challenge (KiTS19) dataset containing 210 labeled and 90 unlabeled patients. On the KiTS19 official website, our dice score for kidney segmentation is 0.9660, and the dice score for tumor segmentation is 0.7407. In this study, we have two contributions. The first point is to improve the performance of 2D models in kidney and tumor segmentation. The second point is to speed up the prediction and reduce the required computing resources. This architecture can make the model flexible for scenarios with limited data or computing resources. In addition, we have established a system with the architecture proposed in this thesis, hoping to make clinical diagnosis more convenient for doctors through this system.

並列關鍵字

Image Segmentation ； Medical Image ； Kidney ； Tumor ； 2D U-Net

參考文獻

[1] M. Shehata, A. Alksas, R. T. Abouelkheir, A. Elmahdy, A. Shaffie, A. Soliman, M. Ghazal, H. A. Khalifeh, A. A. Razek, and A. ElBaz, “A new computer-aided diagnostic (Cad) system for precise identification of renal tumors,” in 2021 IEEE 18th International Symposium on Biomedical Imaging, pp. 1378–1381, 2021.

[2] K. H. Park, K. S. Ryu, and K. H. Ryu, “Determining minimum feature number of classification on clear cell renal cell carcinoma clinical dataset,” in 2016 International Conference on Machine Learning and Cybernetics, vol. 2, pp. 894–898, 2016.

[3] S. H. Raza, Y. Sharma, Q. Chaudry, A. N. Young, and M. D. Wang, “Automated classification of renal cell carcinoma subtypes using scale invariant feature transform,” in 2009 Annual International Conference of the IEEE Engineering in Medicine and Biology Society, pp. 6687–6690, 2009.

[4] F. Han, S. Liao, S. Yuan, R. Wu, Y. Zhao, and Y. Xie, “Explainable prediction of renal cell carcinoma from contrast-enhanced CT images using deep convolutional transfer learning and the shapley additive explanations approach,” in 2021 IEEE International Conference on Image Processing, pp. 3802–3806, 2021.

[5] A. M. OsowskaKurczab, T. Markiewicz, M. Dziekiewicz, and M. Lorent, “Combining texture analysis and deep learning in renal tumour classification task,” in 2020 IEEE 21st International Conference on Computational Problems of Electrical Engineering, pp. 1–4, 2020.