利用深度學習演算法於對比劑乳房攝影影像腫瘤偵測與分類之研究

Detection and Classification for Breast Tumors with Contrast Enhanced Mammography Using Deep Learning Methods

林佩頴(Pei-Ying Lin)

指導教授 ： 黃詠暉

共同指導教授 ： 陳泰賓(Tai-Been Chen)

義守大學/智慧科技學院/資訊工程學系/碩士(2022年)

摘要

研究目的: 對比劑乳房攝影（contrast enhanced mammogram, CEM）已被證明比2D 乳房攝影更靈敏和更具特異性。此研究主要目的為評估深度學習中的腫瘤偵測及腫瘤分類於對比劑乳房攝影的成效。
材料與方法: 這項回溯性研究包括從2012 年1 月至2021 年10 月接受對比劑乳房攝影的女性，符合收案對比劑乳房攝影影像女性共121 位，影像350 張。使用Resnet50 與Darknet53 架構建立YOLO(You Only Look Once)深度學習物件偵測模型，結合ADAM 及SGDM 進行優化。評估模型效能方法包括命中次數(Hit counts)、命中率(Hit rates)與平均精確度 (mean Average Precision, mAP)。分類預訓練CNN 模型包含Alexnet、VGG16、VGG19、ResNet50、ResNet101、InceptionV3、MobilenetV2、Nasnetmobile、Shufflenet、Googlenet 及Densenet201，預訓練CNN主要用於影像特徵萃取再經由SVM(Support Vector Machine)、NB(Naive Bayes)及LR(Logistic Regression)建立分類模型。模型評估包含靈敏度(Sensitivity)、特異性(Specificity)、精確度(Precision)、陰性預測值(Negative Predicted Value, NPV)、準確度(Accuracy)及Kappa 一致性
結果: Darknet 53 搭配優化器 ADAM 時，最高命中率為 97%。Resnet 50 搭配優化器 ADAM ，有最高 mAP 為 82%。影像訓練模型在結合不同特徵層及分類器，以Resnet50 特徵層結合SVM 具最高準確度，Accuracy、Sensitivity、Specificity、Precision、NPV、Kappa 值分別為0.97、0.97、0.96、0.99、0.93、0.92。
結論: 預期利用深度學習演算法來偵測與分類對比劑乳房攝影影像中之腫塊是未來使用於乳房影像評估的有用技術。

關鍵字

對比劑乳房攝影 ； 腫瘤 ； 物件偵測 ； 腫瘤分類

並列摘要

Purpose: Contrast enhanced mammogram (CEM) has been proven to be more sensitive and specific than 2D mammography. The main objective of this study was to evaluate the effectiveness of deep learning for tumor detection and tumor classification in contrast mammography.
Methods and Materials: This retrospective study included women who underwent contrast mammography from January 2012 to October 2021. A total of 121 women with 350 images were eligible for contrast mammography. The Resnet50 and Darknet53 architecture were applied to build You Only Look Once (YOLO)deep learning object detection model with optimizers SGDM (Stochastic Gradient Descent with Momentum) and ADAM (Adaptive Moment Estimation). Evaluation methods include hit counts, hit rates and mean average precision (mAP). Classification pre-trained CNN models include Alexnet, VGG16, VGG19, ResNet50, ResNet101, InceptionV3, MobilenetV2, Nasnetmobile, Shufflenet, Googlenet and Densenet201. Pre-trained CNN is mainly used for image feature extraction and then through SVM (Support Vector Machine), NB (Naive Bayes ) and LR (Logistic Regression) to establish a classification model. Model evaluation includes sensitivity, specificity, precision, negative predictive value ( NPV), accuracy and Kappa consistency
Results: Darknet 53 with the optimizer ADAM has a maximum hit rate of 97%. Resnet 50 with optimizer ADAM has the highest mAP of 82%. The image training model combines different feature layers and classifiers, and the Resnet50 feature layer combined with SVM has the highest accuracy.
Conclusion: The use of deep learning algorithms to detection and classification masses in contrast mammography images is expected to be a useful technique for breast image assessment in the future.

並列關鍵字

CEM ； mass ； YOLO ； Object Detection ； classification

參考文獻

1. Perek, S., et al., Classification of contrast-enhanced spectral mammography (CESM) images. Int J Comput Assist Radiol Surg, 2019. 14(2): p. 249-257.

2. Song, J., et al., Multiview multimodal network for breast cancer diagnosis in contrast-enhanced spectral mammography images. Int J Comput Assist Radiol Surg, 2021. 16(6): p. 979-988.

3. Zhicheng Jiao, X.G., Ying Wang, Jie Li,, A deep feature based framework for breast masses classification,. Neurocomputing,, 2016,. Volume 197,: p. 221-231

4. Agarwal, R., et al., Deep learning for mass detection in Full Field Digital Mammograms. Comput Biol Med, 2020. 121: p. 103774.

5. Ribli, D., et al., Detecting and classifying lesions in mammograms with Deep Learning. Sci Rep, 2018. 8(1): p. 4165.