深度學習在肝臟及腎臟超音波影像分割之研究

Segmentation of Liver and Right Kidney on Ultrasound Images by Fully Convolutional Networks

巫宗軒(Tsung-Hsuan Wu)

指導教授 ： 王祺元

義守大學/智慧科技學院/資訊工程學系/碩士(2022年)

摘要

超音波腹部造影具有非侵入性、無輻射、即時成像、隨處可得、高性價比、方便攜帶等優勢，已成為臨床第一線腹部診斷之工具，但是超音波是極為依賴操作者的造影方式，影像品質易受到造影條件及受檢者條件之影響，使得造影器官紋理及邊界模糊難辨，對於新進醫事人員具有極大之挑戰。近年深度學習廣泛的應用在影像分析及分割，因此本研究嘗試利用深度學習的方法，自動辨識及標示腹部超音波影像上肝、腎邊界。
本研究採用回朔性收集實驗設計，收集臨床腹部造影超音波序列影像588張，影像中包含肝臟(287張)及右肝腎影像(301張)，影像中肝臟及腎臟之區域皆由資深放射師手動標記。分別使用肝臟影像建立肝臟分割模型，右肝腎影像及所有影像建立肝、腎分割模型。影像隨機分配為訓練集及測試集，訓練集用於訓練全卷積網路模型(Fully convolutional networks, FCN)，測試集用於評估所建立模型之泛化能力。模型卷積神經網路(Convolutional Neural Networks, CNN)骨幹，分別採用Xception、Resnet-18、Resnet-50、Mobilenetv2及Inceptionresnetv2等。使用六種驗證指標，包括整體準確度(Global accuracy)、平均準確度(Mean accuracy)、平均交疊率(Mean intersection over union)、加權交疊率(Weighted intersection over union)、平均邊界F-1分數(Mean boundary F-1 score)以及骰子係數(Dice score)，測試並選擇最佳分割模型。
研究結果顯示使用所有影像，Xception卷積神經網路骨幹，搭配ADAM優化器之FCN分割模型，能有效且穩健分割超音波影像上之肝臟及腎臟，其整體準確度、平均準確度、平均交疊率、加權交疊率、平均邊界F-1分數以及骰子係數分別為0.97、0.95、0.89、0.95、0.78及0.94。此外，建立之最佳分割模型應用於不同程度脂肪肝與正常肝臟影像中，分割性能並沒有顯著差異。
研究結論顯示，使用深度學習全卷積網路分割模型，可有效地分割腹部超音波影像中的肝臟和腎臟。儘管超音波回音受到衰減和散射，導致深層器官邊界不明顯，深度學習方法仍有效分割肝腎器官。未來能將此方法應用在超音波影像多器官辨識及相關醫學影像。

關鍵字

超音波 ； 全卷積網路 ； 影像分割

並列摘要

Abdominal ultrasound has the advantages of non-invasiveness, no radiation damage, rapid imaging, low cost, and easy portability when necessary. It has become the first abdominal diagnosis tool in clinical practice. However, image quality is easily affected by operating setup and subject conditions, which makes the texture and boundaries of imaging organs blurred and difficult to distinguish. It is a challenge for inexperienced medical staff. Deep learning has been widely used in image analysis and segmentation in recent years. Hence, this study uses deep learning methods to automatically identify and label the boundaries of the liver and kidney on abdominal ultrasound images.
This study is a retrospective collection experimental design. Five hundred eighty-eight clinical abdominal ultrasound sequences were collected in this study, including 287 liver images and 301 liver with right kidney images. An experienced radiologist labeled the golden of truth with a freehand labeling tool. Images were used to establish ultrasound segmentation models. The images were randomly allocated into a training set and test set. The training set was used to build the Fully Convolutional Network (FCN) models, and the test set was used to evaluate the performance of the built models. The Convolutional Neural Networks (CNN) backbones of the FCN models were Xception, Resnet-18, Resnet-50, Mobilenetv2, and Inceptionresnetv2. Six validation indexes, including Global accuracy, Mean accuracy, Mean intersection over union, Weighted intersection over union, Mean boundary F-1 Score, and Dice score, were used to evaluate performance and select the feasible segmentation model.
In the result, the FCN segmentation model with Xception CNN backbone and the ADAM optimizer can effectively and robustly segment the liver and kidneys on ultrasound images. The Global accuracy, Mean accuracy, Mean intersection over union, Weighted intersection over union, Mean boundary F-1 Score, and Dice score were 0.97, 0.95, 0.89, 0.95, 0.78, and 0.94, respectively. There was no significant difference in segmentation performance between the built mode applied to fatty and normal ultrasound images.
In conclusion, the FCN segmentation model can robustly segment the liver and kidneys in abdominal ultrasound images. Ultrasound beam attenuation and scatter influenced the image quality and resulted in indistinct deep organ boundaries. The segmented performance of the build model was not affected, and segmented liver and kidney robustly. In the future, this method can be applied to multi-organ identification and related medical images.

並列關鍵字

Ultrasound ； Fully Convolutional Networks ； Image segmentation

參考文獻

[1] Ferraioli G, et al. Ultrasound-based techniques for the diagnosis of liver steatosis. World J Gastroenterol. pp. 6053-6062. 2019.

[2] Pirmoazen AM, et al. Quantitative ultrasound approaches for diagnosis and monitoring hepatic steatosis in nonalcoholic fatty liver disease. Theranostics. pp. 4277-4289. 2020.

[3] Chou TH, et al. Deep learning for abdominal ultrasound: A computer-aided diagnostic system for the severity of fatty liver. J Chin Med Assoc. pp. 842-850. 2021.

[4] Fu Y, et al. A review of deep learning based methods for medical image multi-organ segmentation. Phys Med. pp. 107-122. 2021.

[5] Long J, et al. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3431-3440.2015.