融合紋理資訊之深度學習模型判斷乳房X光攝影腫塊良惡性

A Deep Learning Model Integrating Texture Features for Differential Diagnosis of Benign and Malignant Mass in Mammogram

陳柏瑋(Bo-Wei Chen)

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(2020年)

https://doi.org/10.6342/NTU202004356

摘要

乳癌為造成全世界女性死亡的主要癌症之一，透過國內衛生福利部資料顯示:國內女性發生率最高的癌症也同為乳癌，有鑒於乳癌的盛行率，政府利用乳房X光攝影作為乳癌篩檢的初步診斷工具。乳房X光攝影提供了乳房的鈣化點以及腫塊資訊，使無病症的零期乳癌得以被發現，大幅提高患者的存活率。
透過醫學實證，乳癌篩檢確實降低了乳癌的致死率，但龐大的乳房X光攝影影像同時加重了放射科醫師的負擔。且因為乳房組織的複雜性，以及乳房腫塊之間的變異性，促使放射科醫師之間依據經驗的不同，對於乳房腫塊惡性程度的主觀判定存在著觀察者間的差異性(Inter-observer variability)。為了減輕放射科醫師的負擔以及降低觀察者之間的差異性，發展電腦輔助診斷(Computer-aid diagnosis, CAD)，提供醫師客觀乳房腫塊良惡性分類結果就顯得相當重要。
自從Alex Krizhevsky利用卷積神經網路的深度學習模型(Convolutional Neuron Network, CNN)，在數量高達120萬的影像辨識競賽(ILSVRC)中以15.3%錯誤率取得冠軍後，促使近年來針對影像辨識的課題，大多朝向深度學習的方向進行開發。而在乳房攝影腫塊良惡性分類的議題上，以深度學習為導向的CAD開發上，也取得良好的成效。
雖然深度學習達到良好的影像辨識能力，但訓練模型所花費的成本相當可觀，其中樣本數量的多寡更是直接影響模型影像辨識的準確性。而在乳房X光攝影中，由於乳房腫塊的多樣性，訓練樣本不足的問題更是惡化了分類的準確性。為克服樣本不足的問題，本研究開發一學習紋理資訊為導向的深度學習模型。盼望在有限的影像數量下，利用其他深度學習模型所學習的特徵與紋理資訊相互融合，提高乳房腫塊良惡性分類的準確率。
開發學習紋理為導向模型的過程中，Gabor Filter bank具有近似於人類初級視覺皮質層(Primary visual cortex) 擷取紋理資訊的能力，因此本研究發展學習紋理資訊為導向的深度學習模型是以Gabor filter bank為基底。透過第一層卷積層濾波器為Gabor filter bank的設計，可以分別在紋理影像數據集(Kylberg Texture Dataset、Kth-Tips2-b Dataset)得到0.997±0.002、0.993±0.002辨識率。且在這兩個紋理數據集的樣本數量減少至原本的25%時，相較於單一的深度學習模型，合併AlexNet與本研究開發的深度學習模型可提升紋理影像的辨(Kylberg Texture Dataset提升1.6%；Kth-Tips-2-b Dataset 提升4%)。
相較於紋理數據集的資料量，乳房X光攝影數據集的樣本數量明顯不足。因此為降低模型過擬合(Overfitting)現象的產生，在訓練樣本(Training data)方面採取傳統資料擴增(Data augmentation)的方法，提高乳房腫塊樣本的數量以及多樣性。在模型的開發的研究上，則是以GAP(Global average pooling layer)取代傳統全連接層(Fully-connected layer)的設計，大幅降低訓練參數。此外，為使模型學習多樣化的紋理特徵，本研究將先前所使用的紋理影像數據集(Kylberg Texture Dataset、Kth-Tips2-b Dataset)遷徙學習(Transfer learning)至本研究所開發深度學習模型，並利用Squeeze and Excitation Net強化模型學習紋理資訊的能力。最終透過集成式學習(Ensemble learning)的方法結合VGG16、Inception-V3以及本研究開發的深度學習模型，最終得到乳房腫塊良惡性分類的整體準確率0.80±0.03(Specificity: 0.83±0.02，Sensitivity: 0.76±0.06)。
為輔助放射科醫師診斷乳房腫塊良惡性分類的議題上，本研究開發一學習紋理資訊為導向的深度學習模型，並藉由提供紋理資訊與其他深度學習模型相互融合後，提高模型判斷乳房腫塊良惡性的準確率，克服現有乳房腫塊影像有限的瓶頸。

關鍵字

乳房X光攝影 ； 乳房腫塊良惡性分類 ； 深度學習 ； 卷積神經網路 ； Gabor filter bank ； 紋理特徵 ； 集成式學習

並列摘要

Breast cancer is one of main cancer which causes death among females in the world. According to the Ministry of Health and Welfare statistics, breast cancer is the most commonly occurring cancer in woman. As the prevalence of breast cancer, the government promote mammogram as the initial diagnosis facility for breast cancer screening. The mammogram had the ability to detect micro-calcification and breast mass, the results could check carcinoma in situ which no signs or symptoms, and the outcome could elevate the survival rate.
Evidence-based medicine confirmed that mammogram screening can efficiently reduce the mortality rate. However, an enormous amount of images raised radiologists’ burden simultaneously. The radiologist would have inter-observer variability to determine breast lesion category, the probable reasons are the distinct accumulation of experience, breast tissue complexity, and also the type of mass variability. To overcome this challenge, the development of CAD systems is quite an important issue. Because it could provide automatically diagnosis and objective opinions for the radiologist, which could reduce burden and inter-observer variability among radiologist. Since Alex Krizhevsky utilizes deep learning model to achieves a 15.3% error rate and win the ILSVRC challenge championship. The outcome leads to the trend towards deep learning for image recognition topics. With regard to classification mass benign or malignant issue, development CADx based deep learning model also achieved impressive results.
Despite deep learning achieved brilliant results, training model would cost much time and the number of samples which affect model recognition ability directly. Due to the types of breast mass variability, insufficient samples issue would lead the model to have worse classification results. To overcome this challenge, we proposed the deep learning model which has the ability to learn texture features. Combination of different models’ features from another model and our proposed would improve the classification accuracy for mass benign and malignant in the limited medical images.
The process of developing texture features based model, according to research confirm that Gabor filter bank had similar ability about primary visual cortex in the human, which could extract texture features potential. Hence, our proposed model is based on a Gabor filter bank. According to first convolutional layer kernels based on Gabor filter bank, which could achieve 0.997±0.002 and 0.993±0.002 accuracy in the texture image dataset (Kylberg Texture Dataset、Kth-Tips2-b Dataset) separately. In order to check our proposed network limitation, we reduce the texture image samples with 25% descending progressively ratio. Comparison of single model design, combining features from AlexNet and ours would improve texture image recognition accuracy. (Kylberg Texture Dataset improve 1.6% recognition accuracy; Kth-Tips-2-b Dataset improve 4% recognition accuracy)
Mammogram samples are more insufficient than texture image dataset. To avoid the overfitting situation, we use traditional data augmentation method to increase mass samples and variability. For the model design, we use the GAP layer instead of the FC layer to reduce training parameters. Besides, we use previous texture dataset transfer learning to our proposed network, and utilization of SE Net to strength model which capture texture features ability. And use an ensemble learning algorithm to integrate VGG16, Inception-V3 and our proposed network which could achieve mass classification accuracy 0.80±0.03(Specificity: 0.83±0.02, Sensitivity: 0.76±0.06).
To assist radiologist for the breast mass benign and malignant classification issue, we proposed the network which had the ability for texture features learning. And combination features from another network, which could improve accuracy for the mass classification, and overcome the insufficient medical images issue.

並列關鍵字

Mammogram ； Classification of benign and malignant mass ； Deep learning ； CNN ； Gabor filter bank ； Texture features ； Ensemble learning

參考文獻

[1] Momenimovahed, Zohre, and Hamid Salehiniya. "Epidemiological characteristics of and risk factors for breast cancer in the world." Breast Cancer: Targets and Therapy 11 (2019): 151

[2] 108年國人死因統計結果分析-衛生福利部統計處。民國109年6月16日。檢自https://dep.mohw.gov.tw/DOS/cp-4927-54466-113.html

[3] Kalli, Sirishma, et al. "American joint committee on cancer’s staging system for breast cancer: what the radiologist needs to know." Radiographics 38.7 (2018): 1921-1933.

[4] 統計國人罹患乳癌五年存活率(2006-2011)。檢錄自https://www1.cgmh.org.tw/intr/intr5/c6210/breast%20cancer%20stage.html

[5] 衛福部統計2019年國人罹患乳癌人數，2019。檢錄自https://www.hpa.gov.tw/Pages/Detail.aspx?nodeid=591 pid=980