深度學習應用於乳房超音波影像之電腦輔助診斷

Breast Ultrasound Computer-Aided Diagnosis based on Convolutional Neural Network

蔡智晴(Chih-Ching Tsai)

指導教授 ： 張瑞峰

國立臺灣大學/電機資訊學院/生醫電子與資訊學研究所/碩士(2017年)

https://doi.org/10.6342/NTU201703617

摘要

在女性癌症中，乳癌是盛行率最高的癌症，其死亡率也高居第二位，而早期診斷及治療則可大幅降低致死率。乳癌可以透過乳房超音波做診斷，而電腦輔助診斷可以降低人為差異。隨著電腦運算速率的提升，深度學習已被廣泛應用於各個領域，尤其常用於影像辨識。本研究共採用1225個經病理驗證的腫瘤病例，其中包含847個良性病例以及378個惡性病例。本研究並提出以卷積神經網絡中知名的VGG架構為基礎進行簡化的診斷方法，可以在GPU加速的情況下減少15倍的訓練時間。本實驗結果的準確率為84.39%、靈敏性為74.00%、特異性為88.21%及ROC曲線面積為0.91，與傳統的基於紋路的診斷方法沒有顯著差異。而在採用學習遷移（Transfer Learning）之技術後，實驗結果的ROC曲線面積更可高達0.94且具有顯著差異。即便在僅使用原訓練資料的十分之一的情況下，學習遷移仍可達到ROC曲線面積0.89，與傳統方法的0.79比起來有顯著的進步。因此，相較於傳統診斷方式，基於卷積神經網絡的電腦輔助乳癌診斷方法將更為強健。

關鍵字

乳癌 ； 乳房超音波影像 ； 電腦輔助診斷 ； 深度學習 ； 卷積神經網絡

並列摘要

Breast cancer is the most commonly diagnosed cancer and is the second leading cause of cancer death in women. Therefore, early diagnosis leads to early treatment and reduces mortality rates. Breast ultrasound is used to diagnose breast cancer and computer-aided diagnosis (CADx) has been used to decrease inter-observer variation. With the growth of computing power, particularly graphics processing unit (GPU) computing, deep learning has been applied in many different domains and especially in image recognition. In this study, the diagnostic performance was evaluated with 1225 cases with biopsy-proven diagnosis. There were 847 benign cases and 378 malignant cases. The convolutional neural network (CNN) based method proposed in this study, VGG-Lite, is based on the famous VGG network and can be trained 15 times faster with GPU. VGG-Lite produced diagnostic performance with 84.39% in accuracy, 74.00% in sensitivity, 88.21% in specificity and 0.91 in AUC (Area under the Curve of ROC). There was no significant difference compared to the conventional texture analysis method (p-value > 0.05). After applying transfer learning, AUC as high as 0.94 was obtained. When using only 10% of the original training set, the VGG16 architecture achieved an AUC of 0.89 with transfer learning, showing a significant difference compared to the texture analysis method (AUC=0.79, p-value < 0.05). Hence, CNN-based methods are much more robust than conventional methods.

並列關鍵字

Breast cancer ； Breast ultrasound ； Computer-aided diagnosis ； Deep learning ； Convolutional neural network

參考文獻

[1] R. L. Siegel, K. D. Miller, and A. Jemal, "Cancer statistics, 2016," CA: A Cancer Journal for Clinicians, vol. 66, no. 1, pp. 7-30, 2016.

[3] B. Sahiner et al., "Malignant and Benign Breast Masses on 3D US Volumetric Images: Effect of Computer-aided Diagnosis on Radiologist Accuracy," Radiology, vol. 242, no. 3, pp. 716-724, 2007.

[4] S. Singh, J. Maxwell, J. A. Baker, J. L. Nicholas, and J. Y. Lo, "Computer-aided Classification of Breast Masses: Performance and Interobserver Variability of Expert Radiologists versus Residents," Radiology, vol. 258, no. 1, pp. 73-80, 2011.

[5] C.-M. Chen et al., "Breast Lesions on Sonograms: Computer-aided Diagnosis with Nearly Setting-Independent Features and Artificial Neural Networks," Radiology, vol. 226, no. 2, pp. 504-514, 2003.

[6] M. C. Yang et al., "Robust Texture Analysis Using Multi-Resolution Gray-Scale Invariant Features for Breast Sonographic Tumor Diagnosis," IEEE Transactions on Medical Imaging, vol. 32, no. 12, pp. 2262-2273, 2013.