基於U-Net之心臟超音波醫學影像分割用於輔助醫療

Echocardiogram Segmentation to Assist Diagnosis Based on U-Net

劉家銘(Chia-Ming Liu)

指導教授 ： 黃健興

共同指導教授 ： 林義隆(Yih-Lon Lin)

義守大學/智慧科技學院/資訊工程學系/碩士(2021年)

摘要

本論文的目標是要自動化標記心臟積水的超音波影像，為了讓模型能夠達到能夠切出最接近醫師所認定的積水區域，使用了U-Net架構並且替換掉編碼的部分，改成其他預訓練模型以提升效果。本論文在資料集上補上黑邊，使高度和寬度為32的倍數，使資料可以正常適配在模型上，使用IOU和Dice正確率的計算方式去驗證出最合適的模型。本論文最後的驗證集平均mIOU達到70%，相比原始的U-Net所有部位的IOU提升了接近10%，使用產生出來的心臟腔室面積變化趨勢圖，可以輔助醫師醫療上的資訊參考。

關鍵字

深度學習 ； 心包積水 ； 心包填塞 ； 心臟超音波

並列摘要

The goal of this thesis is to automatically mark the ultrasound images of the pericardial effusion. Use U-Net architecture and change encoder part to other pre-trained models. This thesis fills up the image black border to make the side length a multiple of 32, let dataset can use on our model, use IOU and Dice of correct rate calculation methods to verify the most suitable model. The average mIOU of the final validation set of this thesis reached 70%, which is nearly 10% higher than the IOU of all parts of the U-Net. Using the generated heart area trend can assist doctor more information.

並列關鍵字

Deep learing ； Pericardial effusion ； Cardiac tamponade ； U-Net ； Echocardiogram

參考文獻

[1] Liu, Jiamin, et al. "Cascaded coarse-to-fine convolutional neural networks for pericardial effusion localization and segmentation on CT scans." 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018). IEEE, 2018.

[2] Restrepo, C. Santiago, et al. "Imaging findings in cardiac tamponade with emphasis on CT." Radiographics, vol. 27, no. 6, pp. 1595-1610, 2007.

[3] Nayak, Ashwin, David Ouyang, and Euan A. Ashley. "A DEEP LEARNING ALGORITHM ACCURATELY DETECTS PERICARDIAL EFFUSION ON ECHOCARDIOGRAPHY." Journal of the American College of Cardiology, vol. 75, no. 11 Supplement 1, pp. 1563, 2020.

[4] Pérez-Casares, Alejandro, et al. "Echocardiographic evaluation of pericardial effusion and cardiac tamponade." Frontiers in pediatrics 5 pp. 79, 2017.

[5] Veni, Gopalkrishna, et al. "Echocardiography segmentation based on a shape-guided deformable model driven by a fully convolutional network prior." 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018). IEEE, 2018.