基於貝氏定理以及深度學習針對 乳房攝影腫瘤進行 BI-RADS 分類

A Bayesian Approach to BI-RADS Classification of Mammogram Mass with Deep Learning

張漢庭(Joseph Chang)

指導教授 ： 陳中明

國立臺灣大學/工學院/醫學工程學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201900753

摘要

根據全球癌症發病統計，乳癌位居女性癌症發生率第一位，也是在女性癌症中 的主要死因。最近醫療設備的進步以及乳房 X-ray 攝影技術的成熟，乳房中的微小 鈣化點、局部結構變形或腫塊(mass)等大多都能透過 Mammography 來偵測到，以 致越來越多沒有明顯症狀的早期乳癌能被提早發現。根據研究顯示，透過 Mammography 的篩檢來早期發現乳癌能有效降低 20%至 30%的乳癌死亡率。然 而，乳房組織的組成複雜，在不同放射科醫師間容易導致不同的判斷，難以用客觀 的方式評估。因此，為了讓乳房報告檢查有一致性，美國放射科醫學會發展 BI- RADS 書寫報告方式，即所謂的「乳房影像報告暨資料分析系統」，將 X-Ray 攝影 檢查結果分為七類。然而，目前根據 BI-RADS 對乳房病變的評估仍然大多基於定 性的描述以及主觀的判斷，還是具有相當的變異性，導致分級結果有較高的觀察者 間、內的變異。為了提高乳房 X-ray 攝影結果的準確性和一致性，許多研究團隊嘗 試使用機器學習方法來構建電腦輔助診斷 CAD 系統來進行定量的評估。目前 Mammography 相關的研究成果大多限於針對特定病例並且僅將病變區分為良性或 惡性，然而，一個適當的 CAD 系統必須根據臨床程序去設計，提供與放射科醫師 相同方式的輸出結果才能更佳 CAD 輔助醫療診斷的性能。
本研究提出使用貝氏定理的方式克服現有資料不足之問題，並將乳房 X-ray 攝 影中的腫瘤病變分類為 BI-RADS 3 4 5。在所有 BI-RADS 類別中，BI-RADS 3 4 5 的結果尤為重要。正確的判別 BI-RADS 3 4 5 不僅可以幫助早期篩檢和治療，還可 以避免不必要的切片檢查和手術。由於深度學習需要大量的訓練樣本，因此本研究 基於貝氏定理的方式，將已知的結果輸出透過前機率去推理每一類別的事後機率 來確定良惡性信息是否有助於 BI-RADS 的分類。在模型訓練過程中，image augmentation 以及 transfer learning 等常見的方法也被用於幫助避免 overfitting。本 研究針對 VGG16，ResNet50，DenseNet121 和 Inception-V3 等模型進行了測試，以 確定它們在良惡性和 BI-RADS 分類上的表現。對於良惡性腫瘤分類，Inception-V3 的表現優於其他網絡，總體準確度為 0.854，靈敏度為 0.843，特異性為 0.863。對 於 BI-RADS 分類，Inception-V3 的性能也優於其他網絡，總體準確率為 0.622。由於 Inception-V3 的表現在良惡以及 BI-RADS 的分類都優於其他網路，本研究以 Inception-V3 的結構去改良，透過貝氏定理將先驗機率使良惡性的訊息幫助正規化 訓練過程。這方法將BI-RADS分類accuracy 提升了10%，最終總體準確度為0.726， confusion matrix 顯示 BI-RADS 3 的敏感性為 0.701，BI-RADS 為 4:0.761，BI- RADS 為 5:0.717 。最後使用 CAM 顯現模型改良前後的熱度圖能夠看出加入先 驗機率去訓練能有效的改進模型提取重要的特徵像素。為輔助放射科醫師進行 BI-RADS 分類，本研究利用貝氏定理的方式將良惡性 訊息加入模型的訓練，克服現有的資料不足之問題。 由 confusion matrix 以及 CAM 皆顯示使用貝氏定理結合先驗知識可以幫助提高 BI-RADS 分類結果。

關鍵字

乳房攝影良惡性分類 ； 深度學習 ； 卷積神經網路 ； 資料增強

並列摘要

According to Global Cancer Statistics, breast cancer has been the most commonly diagnosed cancer and also the leading cause of cancer death among females. Recent improvements in medical technology and mammography show that early detections of microcalcifications, structural abnormalities and mass can be identified through mammograms. Studies indicate early detection can effectively reduce breast cancer mortality rate by 20 to 30%. However, it is difficult for radiologists to make consistent and objective evaluations. Consequently, in order to standardize image reporting and reduce confusion in breast imaging interpretations among radiologist, the American College of Radiology established a Breast Image Reporting and Data-analyzing System (BI-RADS), classifying lesions into 7 categories. However, current assessment of breast lesions according to BI-RADS remain qualitative and subjective, with substantial inter and intra-reader variability. In order to improve the accuracy and consistency of mammogram results, many studies have been conducted to build computer-aided diagnosis (CAD) systems using machine learning methods. However, developments and performance of latest classification systems are mostly limited to targeting specific cases and only differentiating lesions into benign or malignant, thus a proper suitable CAD system that outputs results according to BI-RADS with the same manner proceeded as radiologists is required to provide a second opinion in clinical settings.
In this study, novel approaches to classify mass lesions into BI-RADS 3 4 5 in mammograms are explored. Among all the BI-RADS categories, consistent results for BI-RADS 3 4 5 is particularly important. Proper reporting of BI-RADS 3 4 5 can not only help early detection and treatment, but also avoid unnecessary biopsies and surgeries. Since deep learning requires large number of training samples, a Bayesian framework has been investigated to see if incorporating malignancy information can help BI-RADS classification performance. Popular techniques such as data augmentation and transfer learning were also used to help avoid overfitting. State-of-the-art models such as VGG16, ResNet50, DenseNet121, and Inception-V3 were tested to determine each of their performance on both malignancy and BI-RADS classifications. For malignancy classification, Inception-V3 outperformed the rest of the networks with an overall accuracy of 0.854, sensitivity of 0.843 and specificity of 0.863. For BI-RADS classification, Inception-V3 also outperformed other networks with an overall accuracy of 0.622. The trained Inception-V3 network was later used as base model and fine-tuned with prior knowledge through a Bayesian framework to regularize the training process with malignancy estimates. This novel approach increased BI-RADS classification performance by 10% with a final overall accuracy of 0.726, the confusion matrix showed sensitivity of BI-RADS 3: 0.701, BI-RADS 4: 0.761 and BI-RADS 5: 0.717. Class activation maps also helped indicate an improvement in localization during prediction. With limited data, the results show that by using a Bayesian approach to incorporate prior- knowledge can help improve the performance of BI-RADS classification.

並列關鍵字

BI-RADS ； Bayesian ； Deep Learning ； Convolutional Neural Network ； Data Augmentation ； Class Activation Map

參考文獻

[1] Bray, Freddie, et al. "Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries." CA: a cancer journal for clinicians 68.6 (2018): 394-424.

[2] Skandan, S. P. (2016). 5 year Overall survival of triple negative breast cancer: A single institution experience.

[3] Tabár, L., Vitak, B., Chen, H. H. T., Yen, M. F., Duffy, S. W., & Smith, R. A. (2001). Beyond randomized controlled trials. Cancer, 91(9), 1724-1731.

[4] Tabar, L., Yen, M. F., Vitak, B., Chen, H. H. T., Smith, R. A., & Duffy, S. W. (2003). Mammography service screening and mortality in breast cancer patients: 20-year follow-up before and after introduction of screening. The Lancet, 361(9367), 1405- 1410.

[5] Nelson, H. D., Fu, R., Cantor, A., Pappas, M., Daeges, M., & Humphrey, L. (2016). Effectiveness of breast cancer screening: systematic review and meta-analysis to update the 2009 US Preventive Services Task Force recommendation. Annals of internal medicine, 164(4), 244-255.