| 摘要: | WHO 2020年全球癌症報告，女性乳癌約占新發癌症病例的11.7%，死亡人數近68.5萬，台灣則每10萬人口78.9人的發生率（0.0789%），是目前成長率最快的癌症。乳房X光攝影是被醫學證實可以有效早期發現乳癌的檢查方法。然而惟有適切的乳房X光攝影擺位才能使乳房組織充分呈現於影像中而不致乳癌被遺漏診斷。根據研究，乳房X光攝影檢查的敏感度可能從擺位通過評估情況下的84.4%下降到擺位失敗的形況下的66.3%；長期以來多數研究與統計也都證實乳房擺位是影響乳房X光攝影檢查成效的關鍵因素。
近年來人工智慧深度學習有了突破性進展，為了減少不正確的乳房攝影擺位可能會導致錯失乳癌的診斷，本研究應用深度學習方法以卷積神經網路模型辨識臨床影像對於MLO與CC視像獨立評估乳房X光攝影擺位的適切性，以減少乳房X光攝影乳癌篩檢的受檢者因攝影擺位的適不切而被召回重照所造成的困擾，及增加放射科醫師對乳房臨床影像診斷乳癌的確定性，進而提升乳房X光攝影乳癌篩檢的成效。
本研究以回溯方式收集台北市立萬芳醫院2017年1月至2020年12月的乳房X光攝影包含有完成RCC、LCC、RMLO、LMLO等4 Views(視像)擺位的乳房X光攝影300例，共1200張影像，由乳房影像放射科醫師及專業放射師依照ACR 1999 Mammography Quality Control Manual的臨床影像評估項目中的乳房X光攝影MLO View 和CC View擺位的標準，對臨床影像進行擺位適切與不適切之分類；分類完成的影像應用遷移學習技術以卷積神經網路 (CNN) 深度學習方法訓練模型，模擬視覺化辨識不同視像乳房影像中乳房組織呈現的充分性的方法，來評估乳房攝影擺位的適切性，經訓練CNN模型後驗證結果，AUC、F1、Precision都高於90%，證實可達成以人工智慧自動評估乳房X光攝影擺位適切與否的目的。
WHO 2020 Global Cancer Report, women's breast cancer cases for about 11.7% of new-haired cancer cases, nearly 6.85 million deaths, the cancer in Taiwan's incidence (0.0789%) of 78.9 people per 100,000 is currently the fastest growth rate. Mammography is currently the most important medical empirical internationally, and an effectively screening method for find breast cancer. However, only adequate mammography positioning can make the breast tissue fully present in the image and prevent breast cancer from being missed. According to the study, the sensitivity of mammography may drop from 84.4% in an adequate positioning to 66.3% in an inadequate positioning; for a long time, most of the studies and statistics have also confirmed breast positioning It is a key factor affecting the effectiveness of mammography.
In recent years, the deep learning of artificial intelligence has made breakthroughs. In order to reduce incorrect mammography positioning, it may lead to the diagnosis of missed breast cancer. In this study, a deep learning approach was used to identify the adequacy of mammograms with a convolutional neural network model to independently assess mammographic positioning of MLO views and CC views to reduce screening mammography risk of breast cancer. It also increases the certainty that radiologists can diagnose breast cancer with mammogram, thereby improving the effectiveness of mammography to screen for breast cancer.
This study retrospectively collected 300 case of mammography including RCC, LCC, RMLO, LMLO 4 views from January 2017 to December 2020 in Taipei WanFang Hospital, with a total of 1200 images. According to the standards mammography positioning of MLO view and CC view in the mammograms evaluation item of ACR 1999 Mammography Quality Control Manual, breast imaging radiologists and radiological technologist classify mammograms as adequate and inadequate for positioning. Then use transfer learning technology to train the model with the convolutional neural network (CNN) deep learning method, as simulate the method of human visually identifying the adequacy of breast tissue presentation in different view of mammograms to evaluate the mammography positioning, after training the CNN models, the results were verified, and the AUC, F1, and Precision were all higher than 90%, which confirmed that the purpose of automatically evaluating the positioning of mammography with artificial intelligence was achieved. |
